Low-altitude aircraft remote sensing image target detection method, device, equipment and medium

Abstract

The application discloses a low-altitude aerial vehicle remote sensing image target detection method and device, equipment and medium, relates to the computer technical field, including in the preset remote sensing scene, using the target low-altitude aerial vehicle to collect remote sensing image;Based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, the multiscale feature extraction is carried out to the collected remote sensing image to be detected, so as to determine the remote sensing image feature processing result;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, the local and global feature fusion is carried out, so as 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, different spatial resolution features are processed, so as to obtain the target detection result of the remote sensing image to be detected.The application improves the target detection performance and scene adaptability of low-altitude remote sensing image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a remote sensing image target detection method, device and equipment of low-altitude aircraft and medium. BACKGROUND

[0002] At present, the particularity of low-altitude remote sensing image brings great challenges to target detection, which makes it different from general natural image processing tasks. First, the target scale in an image varies greatly, second, the targets are densely arranged and oriented in any direction, in addition, it has unique oblique view and rich ground details, which makes target recognition difficult and prone to false positives and false negatives.

[0003] To solve the above problems, in the existing related scheme, the deep learning method based on convolutional neural network becomes the absolute mainstream, although it avoids the adverse effects caused by manual design of features in traditional related schemes, but it is difficult to effectively model the global dependence and long-distance target association in the image, especially the feature extraction ability of small targets and densely arranged targets. And the standard convolution in this scheme has insufficient ability to model the geometric transformation of the target, and the effect is discounted when processing targets oriented in any direction.

[0004] Therefore, how to solve the problems existing in the related prior art, improve the target detection performance and scene adaptability of low-altitude remote sensing image, and improve the accuracy, reliability and robustness of target recognition, is a problem to be solved by those skilled in the art. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a remote sensing image target detection method, device and equipment of low-altitude aircraft, which can solve the problems existing in the related prior art, significantly improve the feature representation ability of multi-scale, small size, densely arranged and arbitrarily oriented targets, improve the target detection performance and scene adaptability of low-altitude remote sensing image, and improve the accuracy, reliability and robustness of target recognition. The specific scheme is as follows:

[0006] In a first aspect, the present application provides a remote sensing image target detection method of low-altitude aircraft, comprising:

[0007] In a preset remote sensing scene, a target low-altitude aircraft is used to collect remote sensing images to determine the remote sensing images to be detected;

[0008] The preset feedforward neural network in the preset lightweight low-altitude image detection model is used to extract and enhance multi-scale features of the to-be-detected remote sensing image, so as to determine a remote sensing image feature processing result.

[0009] Based on the remote sensing image feature processing result and a multi-layer feature fusion component in the preset lightweight low-altitude image detection model, local features and global features are fused to determine a remote sensing image feature fusion result.

[0010] Based on the remote sensing image feature fusion result and a deconvolution upsampling component in the preset lightweight low-altitude image detection model, different spatial resolution features are processed to complete a target detection operation, and a target detection result corresponding to the to-be-detected remote sensing image is obtained.

[0011] Optionally, the preset feedforward neural network in the preset lightweight low-altitude image detection model is used to extract and enhance multi-scale features of the to-be-detected remote sensing image, so as to determine a remote sensing image feature processing result, including:

[0012] The plurality of mobile window attention components in the preset lightweight low-altitude image detection model are used to extract multi-scale features of the to-be-detected remote sensing image, so as to determine a remote sensing image feature processing result.

[0013] Convolution and nonlinear activation operations are performed on the features in the remote sensing image feature processing result, so as to complete a first feature processing result.

[0014] Group convolution is performed on the first feature processing result, and semantic extraction is performed in each sub-space of the group, so as to determine a second feature processing result.

[0015] Convolution is performed on the second feature processing result, and the first feature processing result is combined to determine a third feature processing result.

[0016] Deconvolution is performed on the third feature processing result to determine a first deconvolution result.

[0017] Based on the deconvolution result and a nonlinear activation function, a remote sensing image feature processing result is determined. The remote sensing image feature processing result includes feature maps corresponding to different scales, respectively.

[0018] Optionally, the plurality of mobile window attention components in the preset lightweight low-altitude image detection model are used to extract multi-scale features of the to-be-detected remote sensing image, including:

[0019] The remote sensing image to be detected is divided into a plurality of local windows based on a plurality of moving window attention components in a preset lightweight low-altitude image detection model, so as to determine a division result;

[0020] In the process of adaptive window mapping based on the division result and a convolution layer, local attention analysis is performed in each window in combination with a window moving mechanism to determine a tensor generation result; wherein the window moving mechanism is used to move a window by half along a horizontal direction and a vertical direction respectively; and the tensor generation result includes a plurality of query tensors, key tensors and value tensors;

[0021] The tensor generation result is normalized to determine a remote sensing image feature processing result.

[0022] Optionally, the remote sensing image feature processing result and the plurality of layer feature fusion components in the preset lightweight low-altitude image detection model are used to fuse local features and global features, including:

[0023] Based on the plurality of layer feature fusion components in the preset lightweight low-altitude image detection model, pyramid feature convolution with different convolution kernel sizes and receptive field ranges is performed on each feature map in the remote sensing image feature processing result to determine a feature map processing result;

[0024] For a target feature map as a deep feature map in the remote sensing image feature processing result, a feature enhancement operation is performed based on a spatial attention mechanism, a channel attention mechanism and the feature map processing result to determine a feature enhancement result corresponding to the target feature map;

[0025] Based on the feature map processing result and the feature enhancement result, aggregation is performed to determine a remote sensing image feature fusion result.

[0026] Optionally, the feature enhancement operation based on the spatial attention mechanism, the channel attention mechanism and the feature map processing result for the target feature map as the deep feature map in the remote sensing image feature processing result includes:

[0027] For the target feature map as the deep feature map in the remote sensing image feature processing result, normalization processing is performed to determine a normalized result;

[0028] The normalized result is processed by using the spatial attention mechanism and the channel attention mechanism respectively to determine a first processing result and a second processing result;

[0029] The first processing result and the second processing result are fused to determine a target processing result;

[0030] From the feature map processing result, a target feature map processing result corresponding to the target feature map is acquired;

[0031] Based on the target processing result and the target feature map processing result, fusion is performed to determine a feature enhancement result corresponding to the target feature map.

[0032] Optionally, based on the remote sensing image feature fusion result and a deconvolution upsampling component in the preset lightweight low-altitude image detection model, different spatial resolution feature processing is performed, including:

[0033] Based on the deconvolution upsampling component in the preset lightweight low-altitude image detection model, multi-scale feature in the remote sensing image feature fusion result is processed by residual block processing and cross-layer connection to determine a residual block processing result;

[0034] The residual block processing result is processed by deconvolution to determine a second deconvolution result;

[0035] The feature enhancement result in the remote sensing image feature fusion result is processed by deconvolution to determine a third deconvolution result;

[0036] Based on the second deconvolution result and the third deconvolution result, convolution and normalization processing are performed to determine a target detection result corresponding to the to-be-detected remote sensing image.

[0037] Optionally, the method further includes:

[0038] Acquiring model training data corresponding to the preset lightweight low-altitude image detection model; the model training data includes a plurality of remote sensing images collected by a low-altitude flying vehicle;

[0039] Based on a stage distillation training strategy and the model training data, in different training stages, the preset lightweight low-altitude image detection model is respectively distilled learned with different scales of features to determine the trained preset lightweight low-altitude image detection model.

[0040] In a second aspect, the application provides a remote sensing image target detection device of a low-altitude flying vehicle, including:

[0041] A remote sensing image acquisition module is configured to acquire remote sensing images by using a target low-altitude flying vehicle in a preset remote sensing scene to determine a to-be-detected remote sensing image.

[0042] The feature extraction module is configured to extract and enhance multi-scale features of the remote sensing image to be detected based on a preset feedforward neural network in a preset lightweight low-altitude image detection model, so as to determine a remote sensing image feature processing result; wherein the preset lightweight low-altitude image detection model is a model constructed based on an improved Transformer, and the preset feedforward neural network comprises a plurality of moving window attention components;

[0043] The feature fusion module is configured to fuse local features and global features based on the remote sensing image feature processing result and a multi-layer feature fusion component in the preset lightweight low-altitude image detection model, so as to determine a remote sensing image feature fusion result.

[0044] The detection result determination module is configured to perform different spatial resolution feature processing based on the remote sensing image feature fusion result and a deconvolution upsampling component in the preset lightweight low-altitude image detection model, so as to complete a target detection operation and obtain a target detection result corresponding to the remote sensing image to be detected.

[0045] In a third aspect, the present application provides an electronic device, comprising:

[0046] A memory configured to save a computer program;

[0047] A processor configured to execute the computer program to implement the steps of the low-altitude aerial vehicle remote sensing image target detection method.

[0048] In a fourth aspect, the present application provides a computer readable storage medium configured to save a computer program, wherein the computer program is executed by a processor to implement the steps of the low-altitude aerial vehicle remote sensing image target detection method.

[0049] It can be seen that, in the present application, in a preset remote sensing scene, a target low-altitude flying vehicle is used to collect a remote sensing image to determine a to-be-detected remote sensing image; a preset feedforward neural network in a preset lightweight low-altitude image detection model is used to extract and enhance multi-scale features of the to-be-detected remote sensing image to determine a remote sensing image feature processing result; the preset lightweight low-altitude image detection model is a model constructed based on an improved Transformer, and the preset feedforward neural network includes a plurality of moving window attention components; a multi-layer feature fusion component in the preset lightweight low-altitude image detection model is used to fuse local features and global features based on the remote sensing image feature processing result to determine a remote sensing image feature fusion result; a deconvolution upsampling component in the preset lightweight low-altitude image detection model is used to process features of different spatial resolutions based on the remote sensing image feature fusion result to complete a target detection operation and obtain a target detection result corresponding to the to-be-detected remote sensing image. That is, in the present application, a low-altitude remote sensing image in a preset remote sensing scene is first obtained, then a preset feedforward neural network in a preset lightweight low-altitude image detection model is used to extract and enhance multi-scale features, then a remote sensing image feature fusion result is determined based on the determined remote sensing image feature processing result and a multi-layer feature fusion component in the preset lightweight low-altitude image detection model, and finally a target detection result corresponding to the to-be-detected remote sensing image is obtained based on the remote sensing image feature fusion result and a deconvolution upsampling component in the preset lightweight low-altitude image detection model. In this way, the problems existing in the related art can be solved, the feature representation capability of a target of multi-scale, small size, and dense arrangement and in any direction is significantly improved, the target detection performance and scene adaptability of a low-altitude remote sensing image are improved, and thus the accuracy, reliability, and robustness of target recognition are improved. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on the provided drawings.

[0051] Figure 1 A low-altitude flying vehicle remote sensing image target detection method flowchart is provided for the present application.

[0052] Figure 2 A target detection overall framework structure schematic diagram based on a preset lightweight low-altitude image detection model is provided for the present application.

[0053] Figure 3A structure diagram of a preset feedforward neural network of a preset lightweight low-altitude image detection model provided by the application is shown in the figure.

[0054] Figure 4 A structure diagram of a multi-layer feature fusion component of a preset lightweight low-altitude image detection model provided by the application is shown in the figure.

[0055] Figure 5 A structure diagram of a deconvolution upsampling component of a preset lightweight low-altitude image detection model provided by the application is shown in the figure.

[0056] Figure 6 A structure diagram of a remote sensing image target detection device of a low-altitude aircraft provided by the application is shown in the figure.

[0057] Figure 7 A structure diagram of an electronic device provided by the application is shown in the figure. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0059] In the prior related solutions, the deep learning method based on the convolutional neural network becomes an absolute mainstream. Although the deep learning method based on the convolutional neural network avoids the adverse effects caused by the manual design of features in the traditional related solutions, it is difficult to effectively model the global dependence relationship and the long-distance target correlation in the image, and especially the feature extraction capability for the micro target and the densely arranged target is insufficient. Moreover, the standard convolution in the solution has insufficient modeling capability for the geometric transformation of the target, and the effect is discounted when processing the target in any direction.

[0060] Therefore, the application provides a remote sensing image target detection scheme of a low-altitude aircraft, which can solve the problems in the prior related solutions, significantly improve the feature representation capability for the targets of multiple scales, small sizes, dense arrangements and any direction, improve the target detection performance and scene adaptability of the low-altitude remote sensing image, and thus improve the accuracy, reliability and robustness of target recognition.

[0061] Referring to Figure 1 The embodiments of the application disclose a remote sensing image target detection method of a low-altitude aircraft, which comprises the following steps:

[0062] In step S11, the target low-altitude aircraft is used to collect the remote sensing image in a preset remote sensing scene, so as to determine the remote sensing image to be detected.

[0063] In this embodiment, first, the target low-altitude aircraft is used to collect remote sensing images, and the collected low-altitude remote sensing images are 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, that is, the scene contains rich ground feature details, and the arrangement of the detection target is complex. The detection target can be a pedestrian, a vehicle, a house, etc.

[0064] In step S12, based on the preset feedforward neural network in the preset lightweight low-altitude image detection model, multi-scale feature extraction and enhancement are performed on the remote sensing image to be detected to determine a remote sensing image feature processing result. The preset lightweight low-altitude image detection model is a model constructed based on an improved Transformer, and the preset feedforward neural network includes a plurality of mobile window attention components.

[0065] In this embodiment, after determining the remote sensing image to be detected, first, the preset feedforward neural network in the preset lightweight low-altitude image detection model is used to process the image. Specifically, based on the plurality of mobile 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 a 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 a first feature processing result. Grouped convolution is performed on the first feature processing result, and semantic extraction is performed in each subspace of the grouping to determine a second feature processing result. Convolution is performed on the second feature processing result, and the first feature processing result is combined to determine a third feature processing result. Deconvolution is performed on the third feature processing result to determine a first deconvolution result. Based on the deconvolution result and a nonlinear activation function, a remote sensing image feature processing result is determined. The remote sensing image feature processing result includes feature maps corresponding to different scales, respectively. Moreover, the multi-scale feature extraction using the plurality of mobile window attention components includes: based on the plurality of mobile window attention components in the preset lightweight low-altitude image detection model, the remote sensing image to be detected is divided into a plurality of local windows to determine a division result. In the process of adaptive window mapping based on the division result and a convolution layer, local attention analysis is performed in each window in combination with a window moving mechanism to determine a tensor generation result. The window moving mechanism is used to move a window by half along the horizontal direction and the vertical direction, respectively. The tensor generation result includes a plurality of query tensors, key tensors, and value tensors. The tensor generation result is normalized to determine a remote sensing image feature processing result.

[0066] It can be understood that the preset lightweight low-altitude image detection model is a model constructed based on an improved Transformer, which adopts an encoder-decoder structure and introduces a multi-layer feature fusion module (MFFM) to realize cross-layer connection, so as to effectively improve the transmission and integration ability of multi-scale information. The framework follows the U-Net target detection architecture in overall design, while significantly enhancing the flexibility and scalability of the model through modular design ideas, so that it can adapt to different complexity and scale task requirements. Combined with Figure 2 As shown in the encoder part, three window-moving attention blocks (WABs) based on the improved Transformer structure, i.e., moving window attention components, are embedded. Combined with Figure 3 As shown in the encoder part, three window-moving attention blocks (WABs) based on the improved Transformer structure, i.e., moving window attention components, are embedded. Combined with

[0067] It should be understood that, as shown in Figure 3 The moving window attention component is the core component in the entire framework, and its design goal is to significantly reduce the computational complexity while maintaining high precision of the model. The component first divides the input features into several local windows according to a fixed size, and generates query (Q), key (K) and value (V) tensors through adaptive window mapping operations:

[0068] .

[0069] In the formula, are learnable parameters, denotes convolution or linear transformation, is the input feature. In this process, convolution or linear transformation is used to project the feature, and normalization operation is used to ensure the stability of the numerical distribution and alleviate the problem of gradient dissipation. Then, the component calculates the local self-attention in each window to capture the dependency within the window, and inputs the result and the residual connection into the convolution module to enhance the context modeling ability. The mapped tensor is first normalized to ensure the stability of the numerical distribution and slow down the gradient dissipation. The moving window attention component calculates the local attention :

[0070] .

[0071] wherein, is the channel dimension, M is a selectable mask matrix, denotes the transpose operation on the feature matrix. However, the single window attention is susceptible to the receptive field limitation and is difficult to depict long-range dependencies. To this end, the moving window attention component introduces a window moving mechanism in adjacent layers, which enables the tokens that are not adjacent to enter the same window in the next layer by moving half of the window along the horizontal and vertical directions, respectively, to realize cross-window interaction. This strategy not only effectively expands the receptive field and approaches the modeling effect of global attention, but also greatly reduces the computational and memory overhead. After the attention and convolution are alternately used and normalized, the component simultaneously plays the advantages of attention in global modeling and the strong points of convolution in capturing local details, thereby showing superior adaptability in complex remote sensing images.

[0072] In combination Figure 2 with the above, after the attention calculation is completed, the features further enter the feature enhancement component in the preset feedforward neural network to enhance the representation ability. Unlike the traditional feedforward structure, the convolution forward feedforward neural network, i.e., the preset feedforward neural network, combines convolution, normalization, convolution, residual connection, and deconvolution operation to realize efficient modeling and resolution recovery of multi-scale context information. The process is as follows:

[0073] .

[0074] wherein, denotes nonlinear activation; Z denotes local attention; Conv denotes convolution; DeConv denotes deconvolution. That is, the input features are first linearly transformed and nonlinearly enhanced through convolution and activation operation, and then the channels are decomposed and semantic information is extracted in the subspace using convolution, which not only reduces the parameter quantity and computational overhead, but also strengthens the modeling capability of cross-channel dependencies; the combination of convolution and residual structure further improves the feature expression while maintaining the stability of the gradient and the integrity of the feature flow; finally, the output features gradually recover the spatial resolution through deconvolution, learn adjustable up-sampling weights to supplement the detail information, and improve the nonlinear ability of the final output through the activation function.

[0075] In this way, as a whole, the preset feedforward neural network is used to enhance the features processed by attention, so that the features are more discriminative in the multi-scale context, and the feature maps generated in the decoding stage have both global semantics and local details. The synergistic effect of the moving window attention and the preset feedforward neural network provides strong support for the detection of complex targets in low-altitude remote sensing images: the former efficiently realizes the modeling of local and global dependencies, and the latter maintains strong semantic expression ability while restoring spatial structure and detail information through convolution and upsampling mechanism, thereby showing superior performance in small target recognition and large scene analysis.

[0076] In step S13, 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, the local features and the global features are fused to determine a remote sensing image feature fusion result.

[0077] In this embodiment, after determining the remote sensing image feature processing result, the multi-level features in the result are sent to a multi-level feature fusion component for further processing. The component plays a key role in the entire framework, which can effectively fuse and enhance feature information from different levels. Specifically: based on the multi-level feature fusion component in the preset lightweight low-altitude image detection model, the feature maps in the remote sensing image feature processing result are respectively subjected to pyramid feature convolution with different convolution kernel sizes and receptive field ranges to determine feature map processing results; for a target feature map in the remote sensing image feature processing result as a deep feature map, based on a spatial attention mechanism, a channel attention mechanism and the feature map processing result, a feature enhancement operation is performed to determine a feature enhancement result corresponding to the target feature map; based on the feature map processing result and the feature enhancement result, aggregation is performed to determine a remote sensing image feature fusion result. The feature enhancement operation based on the spatial attention mechanism, the channel attention mechanism and the feature map processing result for the target feature map in the remote sensing image feature processing result as the deep feature map includes: for the target feature map in the remote sensing image feature processing result as the deep feature map, normalization processing is performed to determine a normalization result; the spatial attention mechanism and the channel attention mechanism are respectively used to process the normalization result to determine a first processing result and a second processing result; the first processing result and the second processing result are fused to determine a target processing result; from the feature map processing result, a target feature map processing result corresponding to the target feature map is obtained; based on the target processing result and the target feature map processing result, fusion is performed to determine the feature enhancement result corresponding to the target feature map. That is, for the low-dimensional feature layer, the component adopts pyramid feature convolution operation to fuse the context information under different receptive fields and output multi-scale features; and for the high-dimensional feature layer, the channel attention and spatial attention mechanisms are applied to reduce the parameter amount while maintaining the discriminative ability of the features. Finally, the component effectively aggregates the processed multi-level features. The multi-level feature maps are finally sent to a decoder to output a target detection result. The overall model improves the detection accuracy while considering the computational efficiency and structural clarity, and has good generalization ability and practicality in complex scenarios.

[0078] In combination Figure 4As shown, it is understood that the core goal of the multi-layer feature fusion component is to fully exploit the semantic expression ability of deep features while maintaining the fine details of shallow features, so as to achieve efficient fusion of multi-scale information and improve the target detection and recognition capability of the model in complex remote sensing scenarios. Specifically, the component receives three layers of feature mappings from the encoder: {I1, I2, I3}, where I1 corresponds to shallow features with the highest resolution but the weakest semantic expression, I2 corresponds to middle-layer features that balance medium resolution and medium semantic intensity, and I3 corresponds to deep features with the lowest resolution but the strongest semantics. In order to take into account the advantages of different levels of features during fusion, the component adopts differentiated processing strategies.

[0079] First of all, all input features {I1, I2, I3} will be sent to the pyramid convolution feature processing component to adapt to the expression needs of different scale features. By setting different convolution kernel sizes and receptive field ranges, this module can extract rich feature representations from local edge texture to global structure semantics layer by layer, thereby generating corresponding three-scale outputs . After convolution processing, the shallow features I1 and the middle-layer features I2 obtain outputs and respectively. Since these two types of features have natural advantages in maintaining spatial resolution and detail sensitivity, they are no longer subjected to additional complex processing, but are directly input as part of the multi-layer feature fusion component's output to the decoder for supplementing spatial structure and edge information in the subsequent reconstruction process. In contrast, the processing of deep features I3 is more complex. Although the preliminary output of I3 after convolution processing has strong semantic abstraction ability, it is relatively insufficient in spatial details and is prone to feature redundancy in multi-scale fusion. To solve this problem, the multi-layer feature fusion component further introduces normalization, spatial attention, and channel attention mechanisms based on . The normalization operation first standardizes the feature distribution, reducing the scale inconsistency between different feature channels, providing more stable input for subsequent attention calculation. Then, the spatial attention mechanism learns the correlation between pixel positions to adaptively emphasize target regions and edge regions, thereby compensating for the lack of spatial resolution in deep features. The channel attention mechanism, on the other hand, models the importance differences between different channels to highlight the feature channels that contribute most to target discrimination and suppress redundant or noisy information. After enhancement by these two types of attention mechanisms, deep features have stronger discriminability and detail sensitivity while maintaining semantic abstraction ability.

[0080] Finally, the enhanced features are fused with the preliminary output to form an optimized high-level feature representation. This result is combined with the shallow output and middle layer outputs The multi-layer feature fusion component is combined with an input decoder for subsequent upsampling and feature reconstruction. Through this branch fusion strategy, the multi-layer feature fusion component effectively improves the semantic expression quality of deep layer features while ensuring the integrity of shallow layer details, enabling the model to balance fine edges and global context when processing targets of different scales. Experimental results show that the multi-layer feature fusion component has significant advantages in small target detection of vehicles, buildings, and complex backgrounds, and can improve detection accuracy while maintaining high computational efficiency.

[0081] In step S14, based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, different spatial resolution feature processing is performed to complete the target detection operation and obtain the target detection result corresponding to the remote sensing image to be detected.

[0082] In combination Figure 5 As shown in the figure, 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, the multi-scale features in the remote sensing image feature fusion result are processed by residual block processing and cross-layer connection to determine the residual block processing result; the residual block processing result is deconvoluted to determine a second deconvolution result; the feature enhancement result in the remote sensing image feature fusion result is deconvoluted to determine a third deconvolution result; and 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 realize 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 to ensure that the output features have both semantic expression and detail fidelity through multi-level feature enhancement and resolution recovery.

[0083] It can be understood that, regarding the deconvolution upsampling component, the component first receives low-resolution semantic features from the deep layers of the encoder and combines the shallow layer detail features transferred across layers as input. Before entering the upsampling process, the input features first pass through the residual block processing, using the identity mapping and residual learning mechanism to effectively alleviate the gradient vanishing problem commonly found in deep networks, while maintaining the stability of the features and enhancing the non-linear expression ability. Because the residual structure can preserve the key information of the input and superimpose the detail residual, the network can learn more rich representations and achieve faster convergence and stronger generalization ability during training. After the residual block enhancement, the features enter the deconvolution layer for step-by-step recovery of spatial size. As a learnable upsampling method, deconvolution (transposed convolution) not only can expand the resolution of the feature map, but also can adaptively adjust the upsampling weights through parameter learning, making the recovery result more refined and consistent with the original image structure. Unlike the fixed interpolation method, deconvolution can further strengthen the semantic expression of the features while expanding the spatial resolution, avoiding the loss of detail information during reconstruction. In the illustrated structure, multiple deconvolution layers act on the input features in a step-by-step stacking manner, each layer outputting a higher resolution feature map, which is optimized through convolution and normalization operations. The normalization layer ensures the stability of the features at each scale in terms of numerical distribution, preventing gradient explosion or contraction, and further improving the robustness of the model during training and inference. The skip connection mechanism plays a crucial role in this component. Through the skip connection, the 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 makes up for the inevitable loss of spatial details in the downsampling process, but also provides the decoder with edge and texture information, thereby significantly improving the structural integrity and edge clarity of the output results. The structure in the figure shows that the fused features are input into the residual block again after fusion, which ensures that the fused features can balance global semantics and local details after enhancement, and provides more robust input for subsequent convolution and normalization steps. Finally, all processed features are compressed in channels and integrated through convolution 1x1, and are output as high-resolution feature maps for the decoding stage after normalization.

[0084] That is, the deconvolution upsampling component enhances feature expression capability through residual blocks, gradually recovers resolution through deconvolution, and makes up for the lack of spatial details through cross-layer connection, ultimately achieving efficient mapping from deep semantic features to high-resolution detailed output. In the complex detection task of low-altitude remote sensing images, this component not only can accurately reconstruct small-scale targets such as vehicles and pedestrians, but also can ensure the edge integrity and detail restoration of large-scale targets such as buildings and roads, thereby significantly improving the overall detection performance and scene adaptability.

[0085] In addition, in the model training process of the preset lightweight low-altitude image detection model, due to the complexity of image scale, rich texture details, various target types, large model parameter space, unstable convergence of training, low knowledge transmission efficiency between feature layers, and difficulty in unified optimization of different semantic levels, and the like. Therefore, a new progressive distillation training strategy (Progressive Knowledge Distillation Strategy, PKD strategy for short) is used in the embodiment. The strategy is a neural network training method combining model distillation and stage-by-stage optimization, and the main purpose is to improve the convergence stability, generalization ability and lightweight performance of the deep network in the low-altitude remote sensing image task (such as ground feature segmentation, target detection, scene reconstruction, etc.). The specific training process is as follows: obtaining model training data corresponding to the preset lightweight low-altitude image detection model; the model training data includes a plurality of remote sensing images collected by a low-altitude flying vehicle; based on the progressive distillation training strategy and the model training data, the preset lightweight low-altitude image detection model is subjected to distillation learning of features of different scales in different training stages, to determine the preset lightweight low-altitude image detection model after training.

[0086] It should be understood that, regarding the above specific model training process using the PKD strategy, the entire training process is divided into multiple stages, each stage performs distillation learning on features of different levels, the distillation target and weight of each stage are dynamically adjusted, and the knowledge is gradually migrated from low layer to high layer. In stage 1, low-level features are distilled, the teacher model provides activation mapping of shallow features (such as edges and textures), the student model aligns the low-level feature distribution, learns basic geometry and detail information, and the distillation loss uses MSE (Mean Squared Error) constraint based on feature map. In stage 2, the middle-level features are distilled, and the distillation focus is shifted to the middle-level structural features (shape, spatial layout), an attention guiding mechanism is introduced, and the teacher feature map is weighted before distillation; at the same time, part of the low-level feature constraint is maintained to maintain the consistency of details. In stage 3, high-level semantics are distilled, the focus is on learning high-level semantic representation (semantic feature distribution), KL divergence (Kullback-Leibler Divergence, also known as relative entropy) is used to realize the consistency of the class semantic space, the distillation weight is dynamically adjusted, the low-level constraint is gradually weakened, the high-level semantic guidance is enhanced, and the model is prevented from overfitting to local textures.

[0087] In this way, through the multi-stage distillation process from shallow to deep, the gradual migration and dynamic optimization of knowledge are realized, thereby significantly improving the training stability and feature expression ability of the model in the remote sensing image task. Moreover, in the low-altitude remote sensing image task, the PKD strategy can effectively stabilize the training process, alleviate the gradient shock problem through phased optimization, enhance the feature coordination, promote the collaborative learning of different levels of features, improve the semantic expression ability, and improve the recognition performance in the fine-grained ground objects and complex background, while maintaining the accuracy, realizing parameter compression and efficient inference. The strategy can be widely applied to low-altitude remote sensing image segmentation tasks, and provides a transferable training paradigm for constructing high-performance lightweight remote sensing models.

[0088] Therefore, in the present application, the low-altitude remote sensing image under the preset remote sensing scene is first acquired, then the multi-scale feature extraction and enhancement are performed using the preset feedforward neural network in the preset lightweight low-altitude image detection model, then the remote sensing image feature processing result is determined, and the multi-layer feature fusion component in the preset lightweight low-altitude image detection model is combined to determine the remote sensing image feature fusion result, and then the target detection result corresponding to the to-be-detected remote sensing image is obtained based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model. In this way, the problems existing in the related prior art can be solved, the feature representation ability of multi-scale, small size, and densely arranged and arbitrarily oriented targets can be significantly improved, the target detection performance and scene adaptability of low-altitude remote sensing images can be improved, and thus the accuracy, reliability, and robustness of target recognition can be improved.

[0089] Referring to Figure 6 The embodiment of the present application also discloses a remote sensing image target detection device of a low-altitude flying vehicle, which comprises:

[0090] The remote sensing image acquisition module 11 is configured to acquire a remote sensing image by using a target low-altitude flying vehicle under a preset remote sensing scene, so as to determine a to-be-detected remote sensing image.

[0091] The feature extraction module 12 is configured to extract and enhance multi-scale features of the to-be-detected remote sensing image based on a preset feedforward neural network in a preset lightweight low-altitude image detection model, so as to determine a remote sensing image feature processing result. The preset lightweight low-altitude image detection model is a model constructed based on an improved Transformer, and the preset feedforward neural network comprises a plurality of moving window attention components.

[0092] The feature fusion module 13 is configured to fuse local features and global features based on the remote sensing image feature processing result and a multi-layer feature fusion component in the preset lightweight low-altitude image detection model, so as to determine a remote sensing image feature fusion result.

[0093] The detection result determination module 14 is configured to perform different spatial resolution 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, to complete a target detection operation, and obtain a target detection result corresponding to the remote sensing image to be detected.

[0094] In some specific embodiments, the feature extraction module 12 can be specifically configured to: perform multi-scale feature extraction on the remote sensing image to be detected based on a plurality of mobile window attention components in the preset lightweight low-altitude image detection model, to determine a remote sensing image feature processing result; perform convolution and nonlinear activation operations on features in the remote sensing image feature processing result, to complete a first feature processing result; perform grouped convolution on the first feature processing result, and perform semantic extraction in each sub-space of the grouping, to determine a second feature processing result; perform convolution on the second feature processing result, and combine the first feature processing result to determine a third feature processing result; perform deconvolution on the third feature processing result, to determine a first deconvolution result; determine a remote sensing image feature processing result based on the deconvolution result and a nonlinear activation function; and the remote sensing image feature processing result includes feature maps corresponding to different scales respectively.

[0095] In some specific embodiments, the feature extraction module 12 can be specifically configured to: divide the remote sensing image to be detected into a plurality of local windows based on a plurality of mobile window attention components in the preset lightweight low-altitude image detection model, to determine a division result; in the process of adaptive window mapping based on the division result and a convolution layer, combine a window moving mechanism to perform local attention analysis in each window, to determine a tensor generation result; wherein the window moving mechanism is configured to move a window by half along a horizontal direction and a vertical direction respectively; the tensor generation result includes a plurality of query tensors, key tensors and value tensors; and perform normalization processing on the tensor generation result, to determine a remote sensing image feature processing result.

[0096] In some specific embodiments, the feature fusion module 13 can be specifically configured to: perform pyramid feature convolution with different kernel sizes and receptive field ranges on each of the feature maps in the remote sensing image feature processing result based on a plurality of multi-layer feature fusion components in the preset lightweight low-altitude image detection model, to determine a feature map processing result; perform feature enhancement operations on a target feature map in the remote sensing image feature processing result as a deep feature map based on a spatial attention mechanism, a channel attention mechanism and the feature map processing result, to determine a feature enhancement result corresponding to the target feature map; and aggregate the feature map processing result and the feature enhancement result, to determine a remote sensing image feature fusion result.

[0097] In some embodiments, the feature fusion module 13 can be specifically configured to: perform normalization processing on a target feature map in the remote sensing image feature processing result as a deep feature map to determine a normalization result; process the normalization result using a spatial attention mechanism and a channel attention mechanism respectively to determine a first processing result and a second processing result; fuse the first processing result and the second processing result to determine a target processing result; obtain a 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 a feature enhancement result corresponding to the target feature map.

[0098] In some embodiments, the detection result determination module 14 can be specifically configured to: perform residual block processing and cross-layer connection on multi-scale features in the remote sensing image feature fusion result based on a deconvolution upsampling component in the preset lightweight low-altitude image detection model to determine a residual block processing result; perform deconvolution processing on the residual block processing result to determine a second deconvolution result; perform deconvolution processing on the feature enhancement result in the remote sensing image feature fusion result to determine a third deconvolution result; and perform convolution and normalization processing based on the second deconvolution result and the third deconvolution result to determine a target detection result corresponding to the remote sensing image to be detected.

[0099] In some embodiments, the detection result determination module 14 can be specifically configured to: obtain model training data corresponding to the preset lightweight low-altitude image detection model; the model training data includes a plurality of remote sensing images collected by a low-altitude flying vehicle; and based on a phased distillation training strategy and the model training data, perform distillation learning of features of different scales on the preset lightweight low-altitude image detection model in different training stages to determine the trained preset lightweight low-altitude image detection model.

[0100] Further, the embodiments of the present application also disclose an electronic device, Figure 7 FIG. 1 is a structural diagram of an electronic device 20 according to an exemplary embodiment, and the contents in the figure cannot be considered as any limitation on the use range of the present application.

[0101] Figure 7A structural schematic diagram of an electronic device 20 is provided in the embodiments of the present application. The electronic device 20 can specifically 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 is configured to store a computer program, and the computer program is loaded and executed by the processor 21 to implement the related steps in the low-altitude aircraft remote sensing image target detection method disclosed in any of the foregoing embodiments. In addition, the electronic device 20 in the embodiments of the present application can be specifically an electronic computer.

[0102] In the embodiments of the present application, the power supply 23 is configured to provide working 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 followed by the communication interface 24 can be any communication protocol applicable to the technical solution of the present application, which is not limited specifically herein; the input / output interface 25 is configured to obtain external input data or output data to the outside, and the specific interface type can be selected according to the specific application needs, which is not limited specifically herein.

[0103] In addition, the memory 22 as a carrier for resource storage can be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc., and the resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage mode can be temporary storage or permanent storage.

[0104] The operating system 221 is configured to manage and control each hardware device on the electronic device 20 and the computer program 222, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of completing the low-altitude aircraft remote sensing image target detection method executed by the electronic device 20 disclosed in any of the foregoing embodiments, the computer program 222 can further include a computer program capable of completing other specific work.

[0105] Further, the present application further discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by the processor to implement the low-altitude aircraft remote sensing image target detection method disclosed in the foregoing embodiments. For the specific steps of the method, refer to the corresponding contents disclosed in the foregoing embodiments, which will not be repeated here.

[0106] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. For the same or similar parts between the embodiments, refer to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts refer to the method part.

[0107] Those skilled in the art will further appreciate that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or any combination thereof. To clearly illustrate the interchangeability of hardware and software, various components have been described above generally in terms of their functionality, without referring to the details of their implementation. Whether such functionality is implemented in hardware or software depends on the particular application and design constraints imposed on the overall system. Skilled persons can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

[0108] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in random access memory (RAM), flash memory, read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0109] Finally, it should be noted that the terms "first", "second", and the like, herein do not denote any order, quantity, combination, or importance, but are used to distinguish one element from another, and cannot imply or infer any such relationship or order between the elements. Moreover, the terms "comprise", "comprises", or any other variant thereof are intended to cover non-exclusive inclusions, such that a process, method, article, or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed, or other elements inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0110] The above provides a detailed description of the technical solutions of the present application. The principles and implementation modes of the present application are described by using specific examples. The above description of the examples is only to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed; in view of the above, the content of the specification should not be understood as limiting the present application.

Claims

1. A method for detecting a target in a remote sensing image of a low-altitude flying vehicle, characterized in that, The application relates to a remote sensing image feature processing method and device. In a preset remote sensing scene, a target low-altitude flying vehicle is used to collect a remote sensing image to determine a to-be-detected remote sensing image; Based on a preset feedforward neural network in a preset lightweight low-altitude image detection model, multi-scale feature extraction and enhancement are performed on the to-be-detected remote sensing image to determine a remote sensing image feature processing result; wherein the preset lightweight low-altitude image detection model is a model constructed based on an improved Transformer, and the preset feedforward neural network comprises a plurality of mobile window attention components; Based on the remote sensing image feature processing result and a multi-layer feature fusion component in the preset lightweight low-altitude image detection model, local features and global features are fused to determine a remote sensing image feature fusion result; Based on the remote sensing image feature fusion result and a deconvolution upsampling component in the preset lightweight low-altitude image detection model, different spatial resolution feature processing is performed to complete a target detection operation and obtain a target detection result corresponding to the to-be-detected remote sensing image; Wherein, 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, the local features and the global features are fused, including: Based on the multi-layer feature fusion component in the preset lightweight low-altitude image detection model, pyramid feature convolution with different convolution kernel sizes and receptive field ranges is respectively performed on each feature map in the remote sensing image feature processing result to determine a feature map processing result; For a target feature map in the remote sensing image feature processing result as a deep feature map, a feature enhancement operation is performed based on a spatial attention mechanism, a channel attention mechanism and the feature map processing result to determine a feature enhancement result corresponding to the target feature map; Based on the feature map processing result and the feature enhancement result, aggregation is performed to determine a remote sensing image feature fusion result; The feature enhancement operation for the target feature map in the remote sensing image feature processing result as a deep feature map based on the spatial attention mechanism, the channel attention mechanism and the feature map processing result includes: Normalization processing is performed on the target feature map in the remote sensing image feature processing result as a deep feature map to determine a normalized result; The normalized result is processed by using the spatial attention mechanism and the channel attention mechanism respectively to determine a first processing result and a second processing result; The first processing result and the second processing result are fused to determine a target processing result; From the feature map processing result, a target feature map processing result corresponding to the target feature map is obtained; Based on the target processing result and the target feature map processing result, fusion is performed to determine the feature enhancement result corresponding to the target feature map. 2.The method for detecting a remote sensing image target of a low-altitude flying vehicle according to claim 1, characterized in that, The feature enhancement operation for the target feature map in the remote sensing image feature processing result as a deep feature map based on the spatial attention mechanism, the channel attention mechanism and the feature map processing result includes: Based on the plurality of moving window attention components in the preset lightweight low-altitude image detection model, multi-scale feature extraction is performed on the to-be-detected remote sensing image to determine a remote sensing image feature processing result. Convolution and non-linear activation operations are performed on the features in the remote sensing image feature processing result to complete a first feature processing result. Grouped convolution is performed on the first feature processing result, and semantic extraction is performed in each subspace of the grouping to determine a second feature processing result. Convolution is performed on the second feature processing result, and the first feature processing result is combined to determine a third feature processing result. Deconvolution is performed on the third feature processing result to determine a first deconvolution result. Based on the deconvolution result and a non-linear activation function, a remote sensing image feature processing result is determined; the remote sensing image feature processing result includes feature maps corresponding to different scales respectively. 3.The method of claim 2, wherein, Based on the plurality of moving window attention components in the preset lightweight low-altitude image detection model, multi-scale feature extraction is performed on the to-be-detected remote sensing image to determine a remote sensing image feature processing result. Based on the plurality of moving window attention components in the preset lightweight low-altitude image detection model, the to-be-detected remote sensing image is divided into a plurality of local windows to determine a division result. In the process of adaptive window mapping based on the division result and a convolution layer, local attention analysis is performed in each window in combination with a window moving mechanism to determine a tensor generation result; the window moving mechanism is used to move a window by half along the horizontal direction and the vertical direction respectively; the tensor generation result includes a plurality of query tensors, key tensors, and value tensors; The tensor generation result is normalized to determine a remote sensing image feature processing result. 4.The method of claim 1, wherein, Based on the remote sensing image feature fusion result and the deconvolution upsampling component in the preset lightweight low-altitude image detection model, different spatial resolution feature processing is performed. 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 a residual block processing result. Deconvolution processing is performed on the residual block processing result to determine a second deconvolution result. Deconvolution processing is performed on the feature enhancement result in the remote sensing image feature fusion result to determine a third deconvolution result. Based on the second deconvolution result and the third deconvolution result, convolution and normalization processing are performed to determine a target detection result corresponding to the to-be-detected remote sensing image. 5.The method for detecting a remote sensing image target of a low-altitude flying vehicle according to claim 1, characterized in that, Further comprising: Obtaining model training data corresponding to the preset lightweight low-altitude image detection model; The model training data includes a plurality of remote sensing images collected by a low-altitude flying vehicle; Based on a stage distillation training strategy and the model training data, different scale feature distillation learning is performed on the preset lightweight low-altitude image detection model in different training stages to determine the trained preset lightweight low-altitude image detection model.

6. A low-altitude aerial vehicle remote sensing image target detection device, characterized in that, Comprising: The remote sensing image acquisition module is configured to acquire remote sensing images by using a target low-altitude flying vehicle in a preset remote sensing scene to determine a remote sensing image to be detected. The feature extraction module is configured to extract and enhance multi-scale features of the remote sensing image to be detected based on a preset feedforward neural network in a preset lightweight low-altitude image detection model to determine a remote sensing image feature processing result, wherein the preset lightweight low-altitude image detection model is a model constructed based on an improved Transformer, and the preset feedforward neural network includes a plurality of moving window attention components. The feature fusion module is configured to fuse local features and global features based on the remote sensing image feature processing result and a plurality of multi-layer feature fusion components in the preset lightweight low-altitude image detection model to determine a remote sensing image feature fusion result. The detection result determination module is configured to process features of different spatial resolutions based on the remote sensing image feature fusion result and a deconvolution upsampling component in the preset lightweight low-altitude image detection model to complete a target detection operation and obtain a target detection result corresponding to the remote sensing image to be detected. The feature fusion module is configured to perform pyramid feature convolution of each feature map in the remote sensing image feature processing result with different convolution kernel sizes and receptive field ranges based on the plurality of multi-layer feature fusion components in the preset lightweight low-altitude image detection model to determine a feature map processing result, perform feature enhancement operation on a target feature map in the remote sensing image feature processing result as a deep layer feature map based on a spatial attention mechanism, a channel attention mechanism, and the feature map processing result to determine a feature enhancement result corresponding to the target feature map, and aggregate the feature map processing result and the feature enhancement result to determine a remote sensing image feature fusion result. The feature fusion module is configured to perform normalization processing on a target feature map in the remote sensing image feature processing result as a deep layer feature map to determine a normalized result, perform processing on the normalized result by using a spatial attention mechanism and a channel attention mechanism to determine a first processing result and a second processing result, fuse the first processing result and the second processing result to determine a target processing result, obtain a 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 a feature enhancement result corresponding to the target feature map.

7. An electronic device, comprising: The memory is configured to save a computer program. The processor is configured to execute the computer program to implement the remote sensing image target detection method of the low-altitude flying vehicle according to any one of claims 1 to 5. The memory is configured to save a computer program.

8. A computer-readable storage medium, characterized in that, The processor is configured to execute the computer program to implement the remote sensing image target detection method of the low-altitude flying vehicle according to any one of claims 1 to 5.

Images