Remote sensing image target detection method and device for low-altitude aircraft, equipment and medium

Abstract

The invention discloses a remote sensing image target detection method and device of a low-altitude aircraft, equipment and a medium, and relates to the technical field of computers, and the method comprises the steps: collecting a remote sensing image through a target low-altitude aircraft in a preset remote sensing scene; based on a preset feedforward neural network in a preset lightweight low-altitude image detection model, performing multi-scale feature extraction on the acquired to-be-detected remote sensing image to determine a remote sensing image feature processing result; performing local and global feature fusion based on the remote sensing image feature processing result and a multi-layer feature fusion component in a preset lightweight low-altitude image detection model to determine a remote sensing image feature fusion result; and based on the remote sensing image feature fusion result and a deconvolution up-sampling assembly in a preset lightweight low-altitude image detection model, performing different spatial resolution feature processing to obtain a target detection result of the to-be-detected remote sensing image. According to the invention, the target detection performance and scene adaptability of the low-altitude remote sensing image are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, equipment and medium for remote sensing image target detection of low-altitude aircraft. Background Technology

[0002] Currently, the unique characteristics of low-altitude remote sensing images pose significant challenges to target detection, distinguishing them from general natural image processing tasks. Firstly, there are extreme variations in target scale, with vastly different target sizes within a single image. Secondly, there are dense and arbitrarily oriented target arrangements. Furthermore, the unique tilted perspective and rich ground details make target recognition difficult and prone to false positives and false negatives.

[0003] To address the aforementioned issues, deep learning methods based on convolutional neural networks have become the dominant approach in existing solutions. While this avoids the negative effects of manually designed features in traditional methods, it struggles to effectively model global dependencies and the association of distant targets, particularly lacking in feature extraction capabilities for small or densely packed targets. Furthermore, the standard convolutional layer in this approach is insufficient in modeling the geometric transformations of targets, resulting in reduced effectiveness when handling targets with arbitrary orientations.

[0004] Therefore, how to solve the problems existing in the current related solutions, improve the target detection performance and scene adaptability of low-altitude remote sensing images, and thus improve the accuracy, reliability and robustness of target recognition, is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for target detection in remote sensing images of low-altitude aircraft. This method addresses the problems existing in current related solutions, significantly improves the feature representation capability for multi-scale, small-sized, densely arranged, and arbitrarily oriented targets, enhances the target detection performance and scene adaptability of low-altitude remote sensing images, and thus improves the accuracy, reliability, and robustness of target recognition. The specific solution is as follows: In a first aspect, this application provides a method for target detection in remote sensing images of low-altitude aircraft, comprising: Under a preset remote sensing scenario, a target low-altitude aircraft is used to acquire remote sensing images in order to identify the remote sensing image to be detected. Based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, multi-scale features are extracted and enhanced on the remote sensing image to be detected to determine the feature processing result of the remote sensing image; wherein, the preset lightweight low-altitude image detection model is a model built based on the improved Transformer, and the preset feedforward neural network includes multiple moving window attention components. Based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, local features and global features are fused to determine the remote sensing image feature fusion result. Based on the feature fusion results of the remote sensing image and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, feature processing at different spatial resolutions is performed to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected.

[0006] Optionally, the step of extracting and enhancing multi-scale features of the remote sensing image to be detected based on the preset feedforward neural network in the preset lightweight low-altitude image detection model to determine the remote sensing image feature processing result includes: Based on multiple moving window attention components in a preset lightweight low-altitude image detection model, multi-scale feature extraction is performed on the remote sensing image to be detected to determine the remote sensing image feature processing result. The features in the remote sensing image feature processing result are subjected to convolution and nonlinear activation operations to complete the first feature processing result; The first feature processing result is grouped and convolved, and semantic extraction is performed in each subspace of the group to determine the second feature processing result; The second feature processing result is convolved and combined with the first feature processing result to determine the third feature processing result; The third feature processing result is deconvolved to determine the first deconvolution result; Based on the deconvolution result and the nonlinear activation function, the remote sensing image feature processing result is determined; the remote sensing image feature processing result includes feature maps corresponding to different scales.

[0007] Optionally, the step of performing multi-scale feature extraction on the remote sensing image to be detected based on multiple moving window attention components in a preset lightweight low-altitude image detection model includes: Based on multiple moving window attention components in a preset lightweight low-altitude image detection model, the remote sensing image to be detected is divided into several local windows to determine the division result. During the adaptive window mapping process based on the partitioning results and convolutional layers, a window moving mechanism is combined to perform local attention analysis in each window to determine the tensor generation results; wherein, the window moving mechanism is used to move half a window along the horizontal and vertical directions respectively; the tensor generation results include several query tensors, key tensors and value tensors; The tensor generation results are normalized to determine the remote sensing image feature processing results.

[0008] Optionally, the fusion of local and global features based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model includes: Based on the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, pyramid feature convolution with different kernel sizes and receptive field ranges is performed on each feature map in the remote sensing image feature processing result to determine the feature map processing result. For the target feature map, which is a deep feature map in the remote sensing image feature processing result, feature enhancement operation is performed based on spatial attention mechanism, channel attention mechanism and the feature map processing result to determine the feature enhancement result corresponding to the target feature map; Based on the feature map processing results and the feature enhancement results, they are aggregated to determine the remote sensing image feature fusion result.

[0009] Optionally, the feature enhancement operation performed on the target feature map, which is a deep feature map in the remote sensing image feature processing result, based on spatial attention mechanism, channel attention mechanism, and the feature map processing result, includes: The target feature map, which is used as a deep feature map in the remote sensing image feature processing result, is normalized to determine the normalization result. The normalization result is processed using spatial attention mechanism and channel attention mechanism respectively to determine the first processing result and the second processing result; The first processing result and the second processing result are fused to determine the target processing result; From the feature map processing results, obtain the target feature map processing result corresponding to the target feature map; Based on the target processing result and the target feature map processing result, they are fused to determine the feature enhancement result corresponding to the target feature map.

[0010] Optionally, the step of performing feature processing at different spatial resolutions based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model includes: Based on the deconvolution upsampling component in the preset lightweight low-altitude image detection model, residual block processing and cross-layer connection are performed on the multi-scale features in the remote sensing image feature fusion result to determine the residual block processing result. The residual block processing result is deconvolved to determine the second deconvolution result; The feature enhancement result in the remote sensing image feature fusion result is deconvolved to determine the third deconvolution result; Based on the second deconvolution result and the third deconvolution result, convolution and normalization processing are performed to determine the target detection result corresponding to the remote sensing image to be detected.

[0011] Optionally, the method further includes: Acquire model training data corresponding to the preset lightweight low-altitude image detection model; the model training data includes several remote sensing images collected by the low-altitude aircraft. Based on the phased distillation training strategy and the model training data, the preset lightweight low-altitude image detection model is subjected to feature distillation learning at different scales at different training stages to determine the preset lightweight low-altitude image detection model after training.

[0012] Secondly, this application provides a remote sensing image target detection device for low-altitude aircraft, comprising: The remote sensing image acquisition module is used to acquire remote sensing images using a target low-altitude aircraft under a preset remote sensing scenario in order to identify the remote sensing image to be detected. The feature extraction module is used to extract and enhance multi-scale features of the remote sensing image to be detected based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, so as to determine the feature processing result of the remote sensing image; wherein, the preset lightweight low-altitude image detection model is a model built based on the improved Transformer, and the preset feedforward neural network includes multiple moving window attention components. The feature fusion module is used to fuse local and global features based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, so as to determine the remote sensing image feature fusion result. The detection result determination module is used to perform feature processing at different spatial resolutions based on the feature fusion result of the remote sensing image and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, so as to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected.

[0013] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the steps of the aforementioned remote sensing image target detection method for low-altitude aircraft.

[0014] Fourthly, this application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of the aforementioned remote sensing image target detection method for low-altitude aircraft.

[0015] As can be seen, in this application, under a preset remote sensing scenario, a target low-altitude aircraft is used to acquire remote sensing images to determine the remote sensing image to be detected; based on a preset feedforward neural network in a preset lightweight low-altitude image detection model, multi-scale features are extracted and enhanced on the remote sensing image to be detected to determine the remote sensing image feature processing result; wherein, the preset lightweight low-altitude image detection model is a model built based on an improved Transformer, and the preset feedforward neural network includes multiple moving window attention components; based on the remote sensing image feature processing result and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, local features and global features are fused to determine the remote sensing image feature fusion result; based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, feature processing at different spatial resolutions is performed to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected. In other words, this application first acquires low-altitude remote sensing images of a preset remote sensing scenario. Then, it uses a preset feedforward neural network based on a preset lightweight low-altitude image detection model to extract and enhance multi-scale features. Next, using the determined remote sensing image feature processing results and combining them with the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, it determines the remote sensing image feature fusion result. Finally, based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, it obtains the target detection result corresponding to the remote sensing image to be detected. This approach solves the problems existing in related solutions, significantly improves the feature representation ability for multi-scale, small-sized, densely arranged, and arbitrarily oriented targets, enhances the target detection performance and scene adaptability of low-altitude remote sensing images, and thus improves the accuracy, reliability, and robustness of target recognition. 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 A flowchart of a remote sensing image target detection method for low-altitude aircraft provided in this application; Figure 2 A schematic diagram of the overall framework structure for target detection based on a preset lightweight low-altitude image detection model provided in this application; Figure 3 A schematic diagram of the structure of a preset feedforward neural network for a preset lightweight low-altitude image detection model provided in this application; Figure 4 A schematic diagram of the structure of a multi-layer feature fusion component for a preset lightweight low-altitude image detection model provided in this application; Figure 5 A schematic diagram of the structure of a deconvolution upsampling component for a preset lightweight low-altitude image detection model provided in this application; Figure 6 A schematic diagram of a remote sensing image target detection device for a low-altitude aircraft provided in this application; Figure 7 This application provides a structural diagram of an electronic device. 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] Among existing solutions, deep learning methods based on convolutional neural networks have become the dominant approach. While this avoids the negative effects of manually designed features in traditional solutions, it struggles to effectively model global dependencies and distant target associations in images, particularly lacking in feature extraction capabilities for small targets and densely packed targets. Furthermore, the standard convolution in this approach is insufficient in modeling the geometric transformations of targets, resulting in reduced effectiveness when handling targets with arbitrary orientations.

[0020] To address this, this application provides a target detection scheme for remote sensing images of low-altitude aircraft, which can solve the problems existing in the existing related schemes, significantly improve the feature representation ability of multi-scale, small-sized, densely arranged and arbitrarily oriented targets, improve the target detection performance and scene adaptability of low-altitude remote sensing images, and thus improve the accuracy, reliability and robustness of target recognition.

[0021] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for target detection in remote sensing images of low-altitude aircraft, comprising: Step S11: Under the preset remote sensing scenario, use the target low-altitude aircraft to acquire remote sensing images in order to determine the remote sensing image to be detected.

[0022] In this embodiment, a target low-altitude aircraft first acquires remote sensing images, which are then used as the remote sensing images to be detected. It should be understood that the preset remote sensing scene can be a complex remote sensing scene, meaning the scene contains rich details of ground features and the arrangement of the detection targets is complex. The detection targets can be pedestrians, vehicles, buildings, etc.

[0023] Step S12: Based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, extract and enhance multi-scale features of the remote sensing image to be detected to determine the feature processing result of the remote sensing image; wherein, the preset lightweight low-altitude image detection model is a model built based on the improved Transformer, and the preset feedforward neural network includes multiple moving window attention components.

[0024] In this embodiment, after determining the remote sensing image to be detected, the image is first processed using a preset feedforward neural network in a preset lightweight low-altitude image detection model. Specifically: based on multiple moving window attention components in the preset lightweight low-altitude image detection model, multi-scale feature extraction is performed on the remote sensing image to be detected to determine the remote sensing image feature processing result; convolution and nonlinear activation operations are performed on the features in the remote sensing image feature processing result to complete the first feature processing result; group convolution is performed on the first feature processing result, and semantic extraction is performed in each subspace of the group to determine the second feature processing result; convolution is performed on the second feature processing result, and combined with the first feature processing result, a third feature processing result is determined; deconvolution is performed on the third feature processing result to determine the first deconvolution result; and the remote sensing image feature processing result is determined based on the deconvolution result and the nonlinear activation function; the remote sensing image feature processing result includes feature maps corresponding to different scales. Furthermore, the multi-scale feature extraction using multiple moving window attention components includes: dividing the remote sensing image to be detected into several local windows based on multiple moving window attention components in a preset lightweight low-altitude image detection model to determine the division result; during the adaptive window mapping process based on the division result and the convolutional layer, combining a window moving mechanism, performing local attention analysis in each window to determine the tensor generation result; wherein, the window moving mechanism is used to move half a window along the horizontal and vertical directions respectively; the tensor generation result includes several query tensors, key tensors, and value tensors; and normalizing the tensor generation result to determine the remote sensing image feature processing result.

[0025] Understandably, the preset lightweight low-altitude image detection model is based on an improved Transformer model. It employs an encoder-decoder structure and introduces a multi-layer feature fusion module (MFFM) to achieve cross-layer connections, effectively improving the transmission and integration of multi-scale information. This framework follows the U-Net object detection architecture in its overall design, while significantly enhancing the model's flexibility and scalability through a modular design approach, enabling it to adapt to task requirements of varying complexity and scale. Figure 2 As shown, the encoder section embeds three Window-moving Attention Blocks (WABs) based on an improved Transformer structure, i.e., moving window attention components. Combined with... Figure 3 As shown, each moving window attention component enhances the model's ability to perceive local features while reducing computational complexity by introducing a window attention mechanism. These components are embedded into the standard Transformer architecture and, in conjunction with a convolutional forward feedforward neural network, progressively generate three hierarchical feature maps. These feature maps capture image information at different scales, laying the foundation for subsequent multi-scale feature fusion.

[0026] It is important to understand that, in combination Figure 3 As shown, the moving window attention component is the core component of the entire framework, and its design goal is to significantly reduce computational complexity while maintaining high model accuracy. This component first divides the input features into several local windows of a fixed size, and then generates query (Q), key (K), and value (V) tensors respectively through adaptive window mapping operations: .

[0027] In the formula, All are learnable parameters. Represents convolution or linear transformation. The input features are used as input features. In this process, convolution or linear transformations are applied to project the features, combined with normalization to ensure the stability of the numerical distribution and mitigate gradient vanishing. Subsequently, the component computes local self-attention within each window to capture dependencies within the window, and the result is concatenated with the residual before being input into the convolution module to enhance contextual modeling capabilities. The mapped tensor is first normalized to ensure the stability of the numerical distribution and mitigate gradient vanishing. The moving window attention component computes local attention within the window. : .

[0028] In the formula, Where M is the channel dimension, and M is an optional mask matrix. This represents the transpose operation on the feature matrix. However, single-window attention is easily limited by the receptive field and struggles to characterize long-range dependencies. To address this, the moving window attention component introduces a window-moving mechanism in adjacent layers. By moving half a window horizontally and vertically, tokens that were not originally adjacent enter the same window in the next layer, achieving cross-window interaction. This strategy not only effectively expands the receptive field, approaching the modeling effect of global attention, but also significantly reduces computational and memory overhead. Through alternating attention and convolution operations and normalization, this component leverages the advantages of attention in global modeling and the strength of convolution in capturing local details, thus demonstrating superior adaptability in complex remote sensing images.

[0029] Combination Figure 2 As shown, after attention calculation, the features are further fed into the feature enhancement component of the pre-defined feedforward neural network to enhance representation capabilities. Unlike traditional feedforward structures, the convolutional forward feedforward neural network, i.e., the pre-defined feedforward neural network, combines convolution, normalization, convolution, residual connections, and deconvolution operations to achieve efficient modeling and resolution restoration of multi-scale contextual information. The process is as follows: .

[0030] In the formula, denoted by , Z represents nonlinear activation; Z represents local attention; Conv represents convolution; and DeConv represents deconvolution. In other words, the input features are first linearly transformed and nonlinearly enhanced through convolution and activation operations. Then, convolution decomposes the channels and extracts semantic information within the subspace, reducing the number of parameters and computational overhead while strengthening the modeling ability of cross-channel dependencies. The combination of convolution and residual structures further enhances feature expressiveness while maintaining gradient stability and feature flow integrity. Finally, the output features are gradually restored to spatial resolution through deconvolution, learning adjustable upsampling weights to supplement detailed information, and the nonlinearity of the final output is enhanced through activation functions.

[0031] In this way, the overall role of the pre-set feedforward neural network is to enhance the features after attention processing, making them more discriminative in multi-scale contexts, and ensuring that the feature maps generated in the decoding stage simultaneously possess global semantics and local details. The synergistic effect of the moving window attention and the pre-set feedforward neural network provides strong support for complex target detection in low-altitude remote sensing images: the former efficiently realizes the modeling of local and global dependencies, while the latter maintains strong semantic expressive ability while recovering spatial structure and detail information through convolution and upsampling mechanisms, thus demonstrating superior performance in both small target recognition and large scene parsing.

[0032] Step S13: Based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, perform the fusion of local features and global features to determine the remote sensing image feature fusion result.

[0033] In this embodiment, after determining the remote sensing image feature processing results, the multi-level features in the results are sent to a multi-level feature fusion component for further processing. This component plays a crucial role in the entire framework, effectively fusing and enhancing feature information from different levels. Specifically: based on the multi-level feature fusion component in the preset lightweight low-altitude image detection model, pyramid feature convolutions with different kernel sizes and receptive fields are performed on each feature map in the remote sensing image feature processing results to determine the feature map processing results; for the target feature map, which is a deep feature map in the remote sensing image feature processing results, feature enhancement operations are performed based on spatial attention mechanisms, channel attention mechanisms, and the feature map processing results to determine the feature enhancement results corresponding to the target feature map; based on the feature map processing results and the feature enhancement results, aggregation is performed to determine the remote sensing image feature fusion result. The component performs feature enhancement operations on the target feature map, which serves as a deep feature map in the remote sensing image feature processing results, based on spatial attention mechanisms, channel attention mechanisms, and the feature map processing results. This includes: normalizing the target feature map, which serves as a deep feature map, to determine the normalization result; processing the normalization result using spatial attention mechanisms and channel attention mechanisms respectively to determine a first processing result and a second processing result; fusing the first processing result and the second processing result to determine the target processing result; obtaining the target feature map processing result corresponding to the target feature map from the feature map processing results; and fusing the target processing result and the target feature map processing result to determine the feature enhancement result corresponding to the target feature map. In other words, for low-dimensional feature layers, this component uses pyramid feature convolution operations to fuse contextual information from different receptive fields and output multi-scale features; while for high-dimensional feature layers, channel attention and spatial attention mechanisms are applied to maintain the discriminative ability of the features while reducing the number of parameters. Finally, this component effectively aggregates the processed multi-layer features. The multi-layer feature maps are finally fed into the decoder to output the target detection results. The overall model improves detection accuracy while balancing computational efficiency and structural clarity, and has good generalization ability and practicality in complex scenes.

[0034] Combination Figure 4As shown, it's important to understand that the core objective of the multi-layer feature fusion component is to fully leverage the semantic expressive power of deep features while preserving the detailed information of shallow features, thereby achieving efficient fusion of multi-scale information and improving the model's target detection and recognition capabilities in complex remote sensing scenarios. Specifically, this component receives three layers of feature maps from the encoder: {I1, I2, I3}. I1 corresponds to the shallow features with the highest resolution but weakest semantic expression; I2 corresponds to the mid-level features that balance medium resolution and medium semantic strength; and I3 corresponds to the deep features with the lowest resolution but strongest semantic expression. To balance the advantages of features at different levels during the fusion process, this component employs a differentiated processing strategy.

[0035] First, all input features {I1, I2, I3} are fed into a pyramid convolutional feature processing component to adapt to the representation requirements of features at different scales. By setting different convolutional kernel sizes and receptive field ranges, this module can extract rich feature representations layer by layer from local edge textures to global structural semantics, thereby generating corresponding outputs at three scales. The shallow layer feature I1 and the middle layer feature I2, after convolution processing, yield the outputs respectively. and Because these two types of features have a natural advantage in preserving spatial resolution and detail sensitivity, they do not undergo additional complex processing. Instead, they are directly used as part of the output input of the multi-layer feature fusion component to the decoder to supplement spatial structure and edge information in subsequent reconstruction. In contrast, the processing of deep feature I3 is more complex. Although the initial output of I3 after convolution processing... It possesses strong semantic abstraction capabilities, but its spatial details are relatively insufficient, and it is prone to feature redundancy in multi-scale fusion. To address this issue, multi-layer feature fusion components... Building upon this foundation, normalization, spatial attention, and channel attention mechanisms are further introduced. Normalization first standardizes the feature distribution, reducing scale inconsistencies between different feature channels and providing a more stable input for subsequent attention calculations. Then, the spatial attention mechanism adaptively emphasizes target and edge regions by learning the correlations between pixel locations, thus compensating for the spatial resolution limitations of deep features. The channel attention mechanism, on the other hand, models the importance differences between different channels, highlighting the feature channels that contribute most to target discrimination and suppressing redundant or noisy information. After these two attention mechanisms, deep features retain their semantic abstraction capabilities while possessing stronger discriminative power and detail sensitivity.

[0036] Finally, the enhanced features and initial output The data are then fused to form an optimized high-level feature representation. This result is compared with the shallow output. and middle layer output The data is input into the decoder for subsequent upsampling and feature reconstruction. Through this branching fusion strategy, the multi-layer feature fusion component effectively improves the semantic representation quality of deep features while preserving the integrity of shallow details, enabling the model to handle both fine edges and global context when processing targets at different scales. Experimental results show that the multi-layer feature fusion component has significant advantages in detecting vehicles, buildings, and small targets in complex backgrounds, maintaining high computational efficiency while improving detection accuracy.

[0037] Step S14: Based on the feature fusion result of the remote sensing image and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, perform feature processing at different spatial resolutions to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected.

[0038] Combination Figure 5 As shown, in this embodiment, after determining the remote sensing image feature fusion result, the deconvolution upsampling component in the preset lightweight low-altitude image detection model is used for further processing. Specifically: based on the deconvolution upsampling component in the preset lightweight low-altitude image detection model, residual block processing and cross-layer connections are performed on the multi-scale features in the remote sensing image feature fusion result to determine the residual block processing result; deconvolution processing is performed on the residual block processing result to determine the second deconvolution result; deconvolution processing is performed on the feature enhancement result in the remote sensing image feature fusion result to determine the third deconvolution result; based on the second deconvolution result and the third deconvolution result, convolution and normalization processing are performed to determine the target detection result corresponding to the remote sensing image to be detected. It should be understood that the deconvolution upsampling component is a key component of the entire decoding stage, and its main function is to achieve the gradual recovery of low-resolution features and the fine reconstruction of high-resolution feature maps. This component comprehensively utilizes deconvolution operations, residual blocks, and cross-layer connection mechanisms. Through multi-level feature enhancement and resolution restoration, it ensures that the output features possess both semantic expressiveness and detail fidelity.

[0039] Understandably, the deconvolution upsampling component first receives low-resolution semantic features from deep layers of the encoder and combines them with shallow detail features passed across layers as input. Before entering the upsampling process, the input features are first processed through residual blocks. Utilizing identity mapping and residual learning mechanisms, this effectively alleviates the gradient vanishing problem common in deep networks and enhances nonlinear expressive power while maintaining feature stability. Because the residual structure can retain key information from the input and superimpose detailed residuals, the network can learn richer representations and achieve faster convergence and stronger generalization during training. After residual block enhancement, the features enter the deconvolution layer for gradual spatial size recovery. Deconvolution (transposed convolution), as a learnable upsampling method, not only expands the resolution of the feature map but also adaptively adjusts the upsampling weights through parameter learning, making the recovery result more refined and closely matching the original image structure. Unlike fixed interpolation methods, deconvolution can further enhance the semantic expression of features while amplifying spatial resolution, avoiding the loss of detailed information during reconstruction. In the illustrated structure, multiple deconvolutional layers are stacked sequentially to apply to the input features. Each layer outputs a higher-resolution feature map, which is then optimized through convolution and normalization operations. The normalization layer ensures the stability of the numerical distribution of features at each scale, preventing gradient explosion or shrinkage, and further improving the robustness of the model during training and inference. The skip connection mechanism plays a crucial role in this component. Through skip connections, shallow high-resolution features from the encoder are directly introduced into the upsampling path of the decoder and fused with the deconvolution output features. This not only compensates for the spatial details inevitably lost during downsampling but also provides the decoder with supplementary local information such as edges and textures, significantly improving the structural integrity and edge sharpness of the output. The structure in the figure shows that the cross-layer features and the deconvolution results are fused and then input into the residual block for processing. This design ensures that the enhanced fused features can balance global semantics and local details, and provide a more robust input for subsequent convolution and normalization steps. Finally, all processed features are compressed and integrated through 1×1 convolution, and then normalized to output a high-resolution feature map for the decoding stage.

[0040] In other words, this deconvolutional upsampling component enhances feature representation through residual blocks, gradually restores resolution through deconvolution, and compensates for the lack of spatial details using cross-layer connections, ultimately achieving an efficient mapping from deep semantic features to high-resolution detailed output. In complex detection tasks of low-altitude remote sensing images, this component can not only accurately reconstruct small-scale targets such as vehicles and pedestrians, but also ensure the edge integrity and detail restoration of large-scale targets such as buildings and roads, thereby significantly improving overall detection performance and scene adaptability.

[0041] Furthermore, during the training of the preset lightweight low-altitude image detection model, due to the complex image scale, rich texture details, diverse target types, and large model parameter space, training is prone to problems such as unstable convergence, low knowledge transfer efficiency between feature layers, and difficulty in uniformly optimizing features at different semantic levels. To address this, this embodiment employs a novel Progressive Knowledge Distillation Strategy (PKD strategy). This strategy is a neural network training method that combines model distillation with staged optimization, primarily aimed at improving the convergence stability, generalization ability, and lightweight performance of deep networks in low-altitude remote sensing image tasks (such as ground feature segmentation, target detection, and scene reconstruction). The specific training process is as follows: Model training data corresponding to the preset lightweight low-altitude image detection model is acquired; the model training data includes several remote sensing images collected by a low-altitude aircraft; based on the staged distillation training strategy and the model training data, at different training stages, the preset lightweight low-altitude image detection model undergoes feature distillation learning at different scales to determine the trained preset lightweight low-altitude image detection model.

[0042] It's important to understand that the specific model training process using the PKD strategy described above involves dividing the entire training process into multiple stages. Each stage performs distillation learning for features at different levels, and the distillation objectives and weights of each stage are dynamically adjusted to achieve gradual knowledge transfer from low to high levels. Stage 1 distills low-level features. The teacher model provides activation mappings for shallow features (such as edges and textures), and the student model aligns with these low-level feature distributions, learning basic geometric and detailed information. The distillation loss uses a feature map-based MSE (Mean Squared Error) constraint. Stage 2 distills mid-level features, shifting the focus to mid-level structural features (shape and spatial layout). An attention-guided mechanism is introduced, weighting the teacher feature map before distillation; simultaneously, some low-level feature constraints are maintained to preserve detail consistency. Stage 3 distills high-level semantics, focusing on learning high-level semantic representations (semantic feature distributions). KL divergence (Kullback-Leibler Divergence, also known as relative entropy) is used to achieve consistency in the category semantic space. Distillation weights are dynamically adjusted, gradually weakening low-level constraints and enhancing high-level semantic guidance to prevent the model from overfitting to local textures.

[0043] This multi-stage distillation process, progressing from shallow to deep, enables gradual knowledge transfer and dynamic optimization, significantly improving the model's training stability and feature representation capabilities in remote sensing image tasks. Furthermore, in low-altitude remote sensing image tasks, the PKD strategy effectively stabilizes the training process, mitigating gradient oscillations through staged optimization, enhancing feature coordination, promoting collaborative learning of features at different levels, improving semantic representation capabilities, and enhancing recognition performance under fine-grained ground features and complex backgrounds. It achieves parameter compression and efficient inference while maintaining accuracy. This strategy can be widely applied to tasks such as low-altitude remote sensing image segmentation, providing a transferable training paradigm for building high-performance, lightweight remote sensing models.

[0044] Therefore, this application first acquires low-altitude remote sensing images of a preset remote sensing scene. Then, it uses a preset feedforward neural network based on a preset lightweight low-altitude image detection model to extract and enhance multi-scale features. Next, using the determined remote sensing image feature processing results and combining them with the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, it determines the remote sensing image feature fusion result. Finally, based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, it obtains the target detection result corresponding to the remote sensing image to be detected. This approach solves the problems existing in related solutions, significantly improves the feature representation ability for multi-scale, small-sized, densely arranged, and arbitrarily oriented targets, enhances the target detection performance and scene adaptability of low-altitude remote sensing images, and thus improves the accuracy, reliability, and robustness of target recognition.

[0045] See Figure 6 As shown in the illustration, this application also discloses a remote sensing image target detection device for low-altitude aircraft, comprising: The remote sensing image acquisition module 11 is used to acquire remote sensing images using a target low-altitude aircraft under a preset remote sensing scenario in order to determine the remote sensing image to be detected. The feature extraction module 12 is used to extract and enhance multi-scale features of the remote sensing image to be detected based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, so as to determine the feature processing result of the remote sensing image; wherein, the preset lightweight low-altitude image detection model is a model built based on the improved Transformer, and the preset feedforward neural network includes multiple moving window attention components. The feature fusion module 13 is used to fuse local features and global features based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, so as to determine the remote sensing image feature fusion result. The detection result determination module 14 is used to perform feature processing at different spatial resolutions based on the feature fusion result of the remote sensing image and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, so as to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected.

[0046] In some specific embodiments, the feature extraction module 12 can be used to: perform multi-scale feature extraction on the remote sensing image to be detected based on multiple moving window attention components in a preset lightweight low-altitude image detection model to determine the remote sensing image feature processing result; perform convolution and nonlinear activation operations on the features in the remote sensing image feature processing result to complete the first feature processing result; perform grouped convolution on the first feature processing result and perform semantic extraction in each subspace of the group to determine the second feature processing result; perform convolution on the second feature processing result and combine it with the first feature processing result to determine the third feature processing result; perform deconvolution on the third feature processing result to determine the first deconvolution result; and determine the remote sensing image feature processing result based on the deconvolution result and the nonlinear activation function; the remote sensing image feature processing result includes feature maps corresponding to different scales.

[0047] In some specific embodiments, the feature extraction module 12 can be used to: divide the remote sensing image to be detected into several local windows based on multiple moving window attention components in a preset lightweight low-altitude image detection model to determine the division result; during the adaptive window mapping process based on the division result and the convolutional layer, local attention analysis is performed in each window in conjunction with a window moving mechanism to determine the tensor generation result; wherein, the window moving mechanism is used to move half a window along the horizontal and vertical directions respectively; the tensor generation result includes several query tensors, key tensors, and value tensors; and the tensor generation result is normalized to determine the remote sensing image feature processing result.

[0048] In some specific embodiments, the feature fusion module 13 can be used to: perform pyramid feature convolution with different kernel sizes and receptive field ranges on each feature map in the remote sensing image feature processing result based on the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, so as to determine the feature map processing result; perform feature enhancement operation on the target feature map, which is a deep feature map in the remote sensing image feature processing result, based on spatial attention mechanism, channel attention mechanism and the feature map processing result, so as to determine the feature enhancement result corresponding to the target feature map; and perform aggregation based on the feature map processing result and the feature enhancement result to determine the remote sensing image feature fusion result.

[0049] In some specific embodiments, the feature fusion module 13 can be used to: normalize the target feature map, which is a deep feature map in the remote sensing image feature processing result, to determine the normalization result; process the normalization result using spatial attention mechanism and channel attention mechanism respectively to determine the first processing result and the second processing result; fuse the first processing result and the second processing result to determine the target processing result; obtain the target feature map processing result corresponding to the target feature map from the feature map processing result; and fuse the target processing result and the target feature map processing result to determine the feature enhancement result corresponding to the target feature map.

[0050] In some specific embodiments, the detection result determination module 14 can be specifically used to: perform residual block processing and cross-layer connection on the multi-scale features in the remote sensing image feature fusion result based on the deconvolution upsampling component in the preset lightweight low-altitude image detection model, to determine the residual block processing result; perform deconvolution processing on the residual block processing result to determine the second deconvolution result; perform deconvolution processing on the feature enhancement result in the remote sensing image feature fusion result to determine the third deconvolution result; and perform convolution and normalization processing based on the second deconvolution result and the third deconvolution result to determine the target detection result corresponding to the remote sensing image to be detected.

[0051] In some specific embodiments, the detection result determination module 14 can be used to: acquire model training data corresponding to the preset lightweight low-altitude image detection model; the model training data includes several remote sensing images collected by low-altitude aircraft; based on a phased distillation training strategy and the model training data, at different training stages, the preset lightweight low-altitude image detection model is subjected to feature distillation learning at different scales to determine the trained preset lightweight low-altitude image detection model.

[0052] Furthermore, embodiments of this application also disclose an electronic device, Figure 7 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0053] Figure 7This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the remote sensing image target detection method for low-altitude aircraft disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0054] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0055] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0056] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the remote sensing image target detection method for low-altitude aircraft executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0057] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned remote sensing image target detection method for low-altitude aircraft. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0058] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0059] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0060] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0061] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0062] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. 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 this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for target detection in remote sensing images of low-altitude aircraft, characterized in that, include: Under a preset remote sensing scenario, a target low-altitude aircraft is used to acquire remote sensing images in order to identify the remote sensing image to be detected. Based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, multi-scale features are extracted and enhanced on the remote sensing image to be detected to determine the feature processing result of the remote sensing image; wherein, the preset lightweight low-altitude image detection model is a model built based on the improved Transformer, and the preset feedforward neural network includes multiple moving window attention components. Based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, local features and global features are fused to determine the remote sensing image feature fusion result. Based on the feature fusion results of the remote sensing image and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, feature processing at different spatial resolutions is performed to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected.

2. The remote sensing image target detection method for low-altitude aircraft according to claim 1, characterized in that, The preset feedforward neural network in the preset lightweight low-altitude image detection model extracts and enhances multi-scale features of the remote sensing image to be detected, in order to determine the remote sensing image feature processing results, including: Based on multiple moving window attention components in a preset lightweight low-altitude image detection model, multi-scale feature extraction is performed on the remote sensing image to be detected to determine the remote sensing image feature processing result. The features in the remote sensing image feature processing result are subjected to convolution and nonlinear activation operations to complete the first feature processing result; The first feature processing result is grouped and convolved, and semantic extraction is performed in each subspace of the group to determine the second feature processing result; The second feature processing result is convolved and combined with the first feature processing result to determine the third feature processing result; The third feature processing result is deconvolved to determine the first deconvolution result; Based on the deconvolution result and the nonlinear activation function, the remote sensing image feature processing result is determined; the remote sensing image feature processing result includes feature maps corresponding to different scales.

3. The remote sensing image target detection method for low-altitude aircraft according to claim 2, characterized in that, The method, based on multiple moving window attention components in a preset lightweight low-altitude image detection model, performs multi-scale feature extraction on the remote sensing image to be detected, including: Based on multiple moving window attention components in a preset lightweight low-altitude image detection model, the remote sensing image to be detected is divided into several local windows to determine the division result. During the adaptive window mapping process based on the partitioning results and convolutional layers, a window moving mechanism is combined to perform local attention analysis in each window to determine the tensor generation results; wherein, the window moving mechanism is used to move half a window along the horizontal and vertical directions respectively; the tensor generation results include several query tensors, key tensors and value tensors; The tensor generation results are normalized to determine the remote sensing image feature processing results.

4. The remote sensing image target detection method for low-altitude aircraft according to claim 2, characterized in that, The process of fusing local and global features based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model includes: Based on the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, pyramid feature convolution with different kernel sizes and receptive field ranges is performed on each feature map in the remote sensing image feature processing result to determine the feature map processing result. For the target feature map, which is a deep feature map in the remote sensing image feature processing result, feature enhancement operation is performed based on spatial attention mechanism, channel attention mechanism and the feature map processing result to determine the feature enhancement result corresponding to the target feature map; Based on the feature map processing results and the feature enhancement results, they are aggregated to determine the remote sensing image feature fusion result.

5. The remote sensing image target detection method for low-altitude aircraft according to claim 4, characterized in that, The feature enhancement operation, performed on the target feature map (which serves as a deep feature map) in the remote sensing image feature processing result, based on spatial attention mechanism, channel attention mechanism, and the feature map processing result, includes: The target feature map, which is used as a deep feature map in the remote sensing image feature processing result, is normalized to determine the normalization result. The normalization result is processed using spatial attention mechanism and channel attention mechanism respectively to determine the first processing result and the second processing result; The first processing result and the second processing result are fused to determine the target processing result; From the feature map processing results, obtain the target feature map processing result corresponding to the target feature map; Based on the target processing result and the target feature map processing result, they are fused to determine the feature enhancement result corresponding to the target feature map.

6. The remote sensing image target detection method for low-altitude aircraft according to claim 5, characterized in that, The feature processing based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, performing feature processing at different spatial resolutions, includes: Based on the deconvolution upsampling component in the preset lightweight low-altitude image detection model, residual block processing and cross-layer connection are performed on the multi-scale features in the remote sensing image feature fusion result to determine the residual block processing result. The residual block processing result is deconvolved to determine the second deconvolution result; The feature enhancement result in the remote sensing image feature fusion result is deconvolved to determine the third deconvolution result; Based on the second deconvolution result and the third deconvolution result, convolution and normalization processing are performed to determine the target detection result corresponding to the remote sensing image to be detected.

7. The remote sensing image target detection method for low-altitude aircraft according to claim 1, characterized in that, Also includes: Obtain model training data corresponding to the preset lightweight low-altitude image detection model; The model training data includes several remote sensing images collected by a low-altitude aircraft. Based on the phased distillation training strategy and the model training data, the preset lightweight low-altitude image detection model is subjected to feature distillation learning at different scales at different training stages to determine the preset lightweight low-altitude image detection model after training.

8. A remote sensing image target detection device for a low-altitude aircraft, characterized in that, include: The remote sensing image acquisition module is used to acquire remote sensing images using a target low-altitude aircraft under a preset remote sensing scenario in order to identify the remote sensing image to be detected. The feature extraction module is used to extract and enhance multi-scale features of the remote sensing image to be detected based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, so as to determine the feature processing result of the remote sensing image; wherein, the preset lightweight low-altitude image detection model is a model built based on the improved Transformer, and the preset feedforward neural network includes multiple moving window attention components. The feature fusion module is used to fuse local and global features based on the remote sensing image feature processing results and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, so as to determine the remote sensing image feature fusion result. The detection result determination module is used to perform feature processing at different spatial resolutions based on the feature fusion result of the remote sensing image and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, so as to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the remote sensing image target detection method for low-altitude aircraft as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the remote sensing image target detection method for low-altitude aircraft as described in any one of claims 1 to 7.

Images