An aircraft target domain adaptation detection method based on scene feature decoupling

Abstract

The application discloses an airplane target domain adaptation detection method based on scene feature decoupling and belongs to the technical field of remote sensing image intelligent processing. The application effectively solves the problem that the airplane targets of the source domain and the target domain are significantly different and the model generalization capability is insufficient under the conditions of cross-sensor and cross-scene. By introducing orthogonal subspace modeling, the application realizes independent expression and decoupling of scene features in different dimensions, effectively eliminates feature confusion between domains, and further combines a dynamic scene base expansion mechanism, so that the model can generate an expanded base according to new scene features, thereby enhancing the adaptability and robustness to unknown domains. Without relying on target domain samples, the application realizes cross-domain detection and robust recognition of airplane targets and has excellent domain generalization performance and engineering application value under complex remote sensing imaging conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent processing technology for remote sensing images, specifically relating to an aircraft target domain adaptation detection method based on scene feature decoupling. Background Technology

[0002] High-resolution remote sensing images contain rich target information and have broad application prospects in fields such as airport control and aviation monitoring. Among these, aircraft target detection and recognition are crucial components of intelligent remote sensing interpretation systems, and their performance directly reflects the system's intelligence level and application advancement. In recent years, with the rapid development of deep learning technology, the accuracy and efficiency of aircraft target detection in remote sensing images have been significantly improved. However, under complex conditions such as cross-scene and cross-sensor environments, the generalization ability of models remains significantly insufficient. Different imaging payloads, lighting conditions, geographical environments, and changes in observation angles can all lead to significant distribution differences between the source and target domains, making it difficult to directly apply aircraft target detection models trained on the source domain to new target domain scenarios. Models often experience a sharp decline in detection performance when facing unseen domains, hindering the rapid deployment and efficient application of remote sensing interpretation systems under cross-domain conditions.

[0003] To address this issue, researchers have proposed various domain-adaptive object detection methods to mitigate the differences in target distribution across different domains. The core idea is to use techniques such as feature alignment to enable detectors to maintain stable performance across data domains with significant distributional differences. Early domain-adaptive detection methods typically included pseudo-label-based self-training, adversarial learning-based feature alignment, and image-to-image translation-based input adaptation. These methods improved the generalization ability of models in cross-domain detection tasks to some extent, but still had significant limitations: they mostly relied on pseudo-labels or target domain samples for retraining, restricting the rapid deployment and application of detection models in unknown scenarios. Therefore, in recent years, domain difference correction methods based on feature modulation have gradually become a key research direction in cross-domain detection. These methods introduce a set of learnable feature bases into the network and adaptively modulate the features during the inference phase, thereby achieving rapid adaptation without the involvement of the target domain, and exhibiting greater flexibility and scalability in practical remote sensing intelligent interpretation systems.

[0004] While existing feature-modulation-based domain-adaptive target detection methods can achieve a certain degree of adaptive feature alignment during the inference stage, the lack of structural constraints on the scene feature space leads to high coupling between different scene feature components in the basis, resulting in feature confusion between different domains. Furthermore, the feature basis in these methods is typically fixed, making it difficult to cope with new scene distribution changes. In complex remote sensing applications, especially aircraft target detection tasks, changes in ground background, imaging angle, and illumination conditions can cause scene semantic drift, leading to a significant decrease in the robustness and adaptability of the detector in unknown domains. Therefore, it is necessary to design an aircraft target domain-adaptive detection method capable of decoupling and modeling scene features and possessing dynamic adaptive update capabilities. This would enable robust inter-domain transfer and rapid model deployment even without target domain samples, thereby enhancing the detection accuracy and generalization ability of remote sensing intelligent interpretation systems in complex multi-source environments. Summary of the Invention

[0005] To address the problems of severe scene feature spatial coupling, insufficient expressive power, and difficulty in adapting to cross-domain distribution changes in existing feature modulation methods, this invention proposes an aircraft target domain adaptation detection method based on scene feature decoupling. This method introduces scene feature decoupling and dynamic expansion mechanisms on the basis of the classical feature modulation domain adaptation detection framework, which improves the independence and scalability of scene feature expression, enabling the model to achieve multi-domain adaptation and robust detection without relying on target domain samples.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] First, a set of scene feature bases representing different scene semantics is learned on the source domain image to characterize the diverse imaging conditions and ground cover in the source domain. Further, an orthogonal constraint scene feature modeling mechanism is introduced. Independent scene bases are obtained through orthogonal subspace learning, allowing each base to correspond to different scene feature directions. This decouples features from different domains, reduces redundancy, and enhances the expressive power of scene features. Based on this, a dynamic scene base expansion strategy is designed to adaptively generate new scene bases according to changes in the distribution of source domain features, supplementing the deficiencies in the original scene space representation. This ensures the model can cover new domain changes and avoids feature aliasing of static bases. Through this dynamic expansion mechanism, the model can adaptively modulate according to the feature distribution of the input scene during the inference phase, achieving rapid adaptation and robust detection of unknown domains. The specific technical solution is as follows:

[0008] An aircraft target domain adaptation detection method based on scene feature decoupling includes the following steps:

[0009] The first step is to obtain the source domain image and its enhanced domain image as input, extract multi-scale image features using shared weights, and embed a scene decoupled image backbone network after each scale output. The feature representations of the source domain image and the enhanced domain image are modulated and aligned layer by layer to obtain image feature representations with domain adaptability.

[0010] The second step involves encoding the original text corresponding to the source domain image using a text backbone network to obtain text features. The image features and text features obtained in the first step are then input into the scene decoupling feature enhancer, where they are deeply fused through multi-layer self-attention and cross-attention mechanisms. Scene decoupling modules are embedded in each branch after the deep fusion of image features to further enhance the domain invariance of the image features and improve the semantic alignment capability between the image and text.

[0011] The third step is to determine the coverage capability of the existing scene base set based on the matching similarity between the current sample scene features and the current scene base features in each scene decoupling feature enhancer. When a sample with a matching degree lower than the threshold is found, the statistic of the sample is registered as a new scene base direction to realize the dynamic expansion of the scene feature space.

[0012] The fourth step involves fusing image and text features for language-guided query selection, followed by inputting the data into a cross-modal decoder for target recognition and localization. During training, detection loss and contrastive loss are jointly used for optimization, ultimately achieving cross-domain adaptive detection of aircraft targets.

[0013] An aircraft target domain adaptation detection device based on scene feature decoupling includes:

[0014] The decoupling and modulation module takes the source domain image and its enhanced domain image as input, extracts multi-scale image features using shared weights, and embeds a scene decoupling image backbone network after each scale output. It modulates and aligns the feature representations of the source domain image and the enhanced domain image layer by layer to obtain image feature representations with domain adaptability.

[0015] The semantic alignment and fusion module encodes the original text corresponding to the source domain image into text features through a text backbone network. The obtained image features and text features are then input into the scene decoupling feature enhancer, and deep fusion is performed through a multi-layer self-attention and cross-attention mechanism. The scene decoupling module is embedded in each branch after the deep fusion of image features to further enhance the domain invariance of image features and improve the semantic alignment capability of images and text.

[0016] The scene enhancement module, in each scene decoupling feature enhancer, judges the coverage capability of the existing scene base set based on the matching similarity between the current sample scene features and the current scene base features; when a sample with a matching degree lower than the threshold is found, the statistic of the sample is registered as a new scene base direction, realizing the dynamic expansion of the scene feature space;

[0017] The detection result generation module fuses image and text features for language-guided query selection, and then inputs the results into a cross-modal decoder for target recognition and localization. During training, detection loss and contrast loss are jointly used for optimization, ultimately achieving cross-domain adaptive detection of aircraft targets.

[0018] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0019] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0020] The present invention has the following beneficial effects:

[0021] This invention effectively enhances the expressive power of scene features by constructing a scene feature decoupling mechanism, solving the cross-domain performance degradation problem caused by scene spatial coupling in existing feature modulation methods. By introducing a dynamic expansion strategy, it enhances the scalability of the scene substrate, enabling the model to flexibly cope with distribution changes under multi-source imaging conditions. It can still achieve stable detection under the condition of no target domain samples, significantly improving the generalization ability and system deployment efficiency of remote sensing aircraft target detection in complex cross-domain scenarios. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of an aircraft target domain adaptation detection method based on scene feature decoupling according to the present invention.

[0023] Figure 2 This is a diagram of the network framework proposed in this invention;

[0024] Figure 3 This is a schematic diagram of the embedding of the scene decoupling module; where (a) is the scene decoupling image backbone network and (b) is the scene decoupling feature enhancer.

[0025] Figure 4 This is a structural diagram of the scene decoupling module. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0027] like Figure 1 As shown, the aircraft target domain adaptation detection method based on scene feature decoupling of the present invention includes the following steps:

[0028] The first step is to obtain the source domain image and its enhanced domain image as input, extract multi-scale image features using shared weights, and embed a scene decoupled image backbone network after each scale output. The feature representations of the source domain image and the enhanced domain image are modulated and aligned layer by layer to obtain image feature representations with domain adaptability.

[0029] The second step involves encoding the original text corresponding to the source domain image using a text backbone network to obtain text features. The image features and text features obtained in the first step are then input into the scene decoupling feature enhancer, where they are deeply fused through multi-layer self-attention and cross-attention mechanisms. Scene decoupling modules are embedded in each branch after the deep fusion of image features to further enhance the domain invariance of the image features and improve the semantic alignment capability between the image and text.

[0030] The third step is to determine the coverage capability of the existing scene base set in each scene decoupling feature enhancer based on the matching similarity between the current sample scene features and the current scene base features. When a sample with a matching degree lower than the threshold is found, the statistic of the sample is automatically registered as a new scene base direction, thereby realizing the dynamic expansion of the scene feature space and enhancing the model's adaptability to new scenes.

[0031] The fourth step involves fusing image and text features for language-guided query selection, followed by inputting the data into a cross-modal decoder for target recognition and localization. During training, detection loss and contrastive loss are jointly employed for optimization, guiding the model to learn feature representations with cross-domain generalization capabilities, ultimately achieving cross-domain adaptive detection of aircraft targets.

[0032] Specifically, the overall network framework constructed by this invention is as follows: Figure 2As shown, the system consists of five key modules: a scene-decoupled image branch, a raw text branch, a scene-decoupled feature enhancer, a semantically guided query selection, and a cross-modal decoder. The scene-decoupled image branch learns an orthogonal subspace of scene features through a scene-decoupled image backbone network, performs domain feature correction, and ultimately extracts domain-adaptive features of the image in both the source and enhanced domains. The text branch receives the raw text query, "airplane," and generates corresponding semantic features through a text backbone network. In the subsequent scene-decoupled feature enhancer, similar to the scene-decoupled image backbone network, a cross-domain scene decoupling method is further utilized to simultaneously obtain enhanced text and image domain-adaptive features. Finally, guided by the text semantic features, adaptive query initialization is achieved, followed by the fusion of text and image information and target localization through the cross-modal decoder. The overall training process is optimized through the collaborative driving of detection loss and contrastive loss.

[0033] Figure 3 (a) illustrates a schematic diagram of the scene decoupling image backbone network of the present invention, i.e., the embedding of the scene decoupling module. First, for each input source domain image... , As a high-altitude feature, For width features, first construct the corresponding augmented domain image. These enhancement operations include, but are not limited to, brightness perturbation, color transformation, and contrast adjustment, aiming to expand scene diversity beyond the image distribution during the training phase. Furthermore, the source domain image and the enhanced domain image are jointly input into a scene-decoupled image backbone network with shared weights, extracting multi-scale features layer by layer. To improve the model's scene adaptability at different semantic levels, a scene decoupling module is inserted after each scale output to achieve layer-by-layer feature alignment and domain feature correction. Specifically, for the ... Layer feature extraction operation The feature representations of the source and enhancement domain images at this scale are obtained respectively:

[0034] (1)

[0035] Subsequently, a scenario decoupling module was introduced. Feature modulation for scene decoupling of original features:

[0036] (2)

[0037] The modulation features of all layers will be uniformly fused to form the final output:

[0038] (3)

[0039] in, Represents a linear transformation for channel dimension alignment. This indicates a multi-scale splicing operation. The number of pixels representing image features. This represents the number of channels.

[0040] Meanwhile, corresponding to the image branch, the text branch receives the category name text "airplane" as the query input and directly extracts its vector representation in the semantic space. To further achieve semantically consistent cross-modal fusion and mitigate the interference of inter-domain differences in image features on image-text fusion, this invention further... and The scene decoupling feature enhancer is fed into the scene decoupling feature enhancer, which introduces the scene decoupling module again in each layer of interaction, such as... Figure 3 As shown in (b), this module can improve the cross-domain robustness and semantic matching ability of visual features, and promote the enhancement of image and text features.

[0041] Specifically, let the first The image and text features input to the layer are respectively , , The number of pixels representing text features is used to capture their structural information through self-attention layers. A bidirectional cross-attention mechanism is then executed to obtain the interacted image and text features. This mechanism includes image-to-text cross-attention and text-to-image cross-attention. These features are further enhanced with feedforward neural networks to obtain... and To mitigate domain bias interference, a scene decoupling module is introduced after the output of the feedforward neural network in the image branch. The final output is obtained as follows:

[0042] (4)

[0043] in, This represents a feedforward neural network, with updated features. It is then sent to the next layer of scene decoupling feature enhancer for further image-text fusion enhancement.

[0044] After completing multi-level image and text feature interaction, the final fused image features Text features Used to initialize cross-modal query vectors , as input to the cross-modal decoder. The system performs multi-layer interactive updates along with image and text features, progressively generating object detection results. Similar to the scene decoupling feature enhancer, each layer includes cross-attention, self-attention, and feedforward networks between the query and image / text features. The final output is the target query. Semantic matching is performed with text features to obtain the final category label. To improve the consistency between image and text representations, this invention employs contrastive loss. As a classification loss, each target query is constrained. The text is close to the corresponding category text in the feature space, while being far away from other categories, as shown below:

[0045] (5)

[0046] in, It is the text embedding corresponding to the true category of the target. It represents the number of categories.

[0047] In addition, the model also predicts the target location through a bounding box regression branch, i.e., outputting the predicted box. And the regression loss was calculated. As a detection loss . A smoothed L1 loss was used, which is mainly used to measure... Prediction boxes With real frame The positional differences between them, including the x-axis y-axis ,Width ,high and rotation angle Calculated only for positive foreground samples:

[0048] (6)

[0049] in,

[0050] (7)

[0051] Finally, the detection loss and contrastive loss are jointly optimized to form the complete detection training objective function:

[0052] (8)

[0053] In the above process, the scene decoupling module The specific structure is as follows Figure 4 As shown, its core idea is to deconstruct scene features into a set of independent orthogonal basis vectors, combine scene basis vectors in these orthogonal directions, and finally correct the original scene features through normalization modulation. This independence promotes the accurate characterization of inter-domain differences, avoids the limited scene expressive power caused by feature confusion in various dimensions, and enables the model to achieve adaptive modulation using existing basis vectors when facing an unseen target domain, thereby improving the domain invariance of scene features.

[0054] Specifically, this involves decoupling image features in the image backbone network based on scene context. For example, firstly, by using the mean and variance of the channel dimension, the overall scene statistics are extracted to achieve scene feature extraction:

[0055] (9)

[0056] in, , where represent the channel-level scene mean and variance for each sample, respectively.

[0057] Furthermore, this invention predefines a scene decoupling feature library. ,in, It indicates the first The mean and variance of each basis of a scene can represent an independent feature direction. Therefore, the mean vector in these basis vectors... Orthogonal constraints need to be maintained during training. This spans a decoupled orthogonal subspace. :

[0058] (10)

[0059] in, It is the identity matrix. It is the Frobenius norm. Then, based on the original scene features... and It performs scene feature matching with a predefined scene base and calculates the distance between them. The features used for modulation are obtained. To better measure distribution distance, this invention uses Wasserstein distance for calculation, and the scene features are compared with the first... The scene base in the first The distance in dimension is defined as follows:

[0060] (11)

[0061] Therefore, scene features and the first The similarity score of each scene basis can be represented by the normalized distance as:

[0062] (12)

[0063] Then, the softmax function is used to map the weights to scene matching.

[0064] (13)

[0065] Based on matching weight By performing weighted fusion on all scene bases, we can obtain the feature modulation target statistics of the current scene, i.e., the mixture mean. and mixed variance :

[0066] (14)

[0067] Finally, adaptive instance normalization, i.e., the AdaIN operation, is used to align the image features, obtaining the modulated image domain adaptive features. : (15)

[0068] The modulation mechanism in the scene decoupling feature enhancer is consistent with the above process. Through multiple corrections, it achieves effective alignment between the source domain image and scene features in different domains, reducing the impact of inter-domain differences on the detector.

[0069] Although the above process builds the model's scene decoupling capability and modulates features through a soft matching mechanism, resulting in stronger expressive power, it is limited by the distribution of the initial basis. When the distance between the input image and all existing scene basis points is large, the original basis points struggle to provide effective feature alignment support, making the model less robust in application. Intuitively, a direct alternative is to pre-define a large number of scene basis points, i.e., set... The base case should be large enough to cover as many domain scenarios as possible. However, a larger base case is not always better; redundant base cases not only increase computation but can also lead to issues with matching weights. The modulation is too uniform and unstable. On the other hand, the preset base cannot adapt to changes in the data and cannot effectively capture rare or new combination scenes that gradually emerge during training, thus weakening the robust learning ability of scene features.

[0070] Therefore, this invention further designs a dynamic expansion mechanism for the scene basis, enabling the feature library to adaptively expand new scene basis directions based on the coverage of the real sample distribution during training. When it is detected that the scene features of the current sample have a low matching degree in the existing subspace, the model will automatically generate and introduce a new basis vector to supplement the expressive power of the existing space and continuously enhance the model's ability to eliminate cross-domain feature differences.

[0071] Specifically, the similarity score obtained according to equation (12) The average of the channel dimensions can be used to obtain the overall similarity score, and finally the maximum similarity score can be selected. Used to evaluate whether the existing scene baseline can cover the current sample:

[0072] (16)

[0073] when Below a given threshold When this occurs, it indicates that the scene features of the current image are significantly different from any existing basis, and the model registers the current scene feature representation as a new scene basis. :

[0074] (17)

[0075] After the new basis obtained from the expansion is added to the feature library, it will participate in orthogonal subspace modeling together with other basis in the next training iteration, and obey the orthogonal constraint objective of Equation (10). In fact, this operation is equivalent to constructing a new orthogonal direction for the current scene.

[0076] In summary, the feature modulation method proposed in this invention, which integrates decoupled orthogonal subspace modeling and dynamic expansion mechanisms, effectively achieves cross-domain scene decoupling capability, enhancing the model's adaptability and modulation stability to multi-domain features in complex remote sensing images. Specifically, orthogonal scene decoupling ensures the independence and low redundancy of each feature dimension in the scene feature space, making the modulation process more discriminative and controllable. The dynamic expansion mechanism, on the other hand, endows the model with the ability to autonomously perceive scene deficiencies and continuously introduce new basis during training, thereby constantly enriching the scene representation space and improving the model's robustness.

[0077] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An aircraft target domain adaptation detection method based on scene feature decoupling, characterized in that, Includes the following steps: The first step is to obtain the source domain image and its enhanced domain image as input, extract multi-scale image features using shared weights, and embed a scene decoupled image backbone network after each scale output. The feature representations of the source domain image and the enhanced domain image are modulated and aligned layer by layer to obtain image feature representations with domain adaptability. The second step involves encoding the original text corresponding to the source domain image using a text backbone network to obtain text features. The image features and text features obtained in the first step are then input into the scene decoupling feature enhancer, where they are deeply fused through multi-layer self-attention and cross-attention mechanisms. Scene decoupling modules are embedded in each branch after the deep fusion of image features to further enhance the domain invariance of the image features and improve the semantic alignment capability between the image and text. The third step is to determine the coverage capability of the existing scene base set based on the matching similarity between the current sample scene features and the current scene base features in each scene decoupling feature enhancer. When a sample with a matching degree lower than the threshold is found, the statistic of the sample is registered as a new scene base direction to realize the dynamic expansion of the scene feature space. The fourth step involves fusing image and text features for language-guided query selection, followed by inputting the data into a cross-modal decoder for target recognition and localization. During training, detection loss and contrastive loss are jointly used for optimization, ultimately achieving cross-domain adaptive detection of aircraft targets. The scene decoupling image backbone network in the first step includes a pluggable scene decoupling module for learning an orthogonal scene feature basis that constitutes the scene space. The scene decoupling module specifically includes: The scene features of each sample are constructed using the channel mean and variance of image features as scene statistics. Multiple learnable orthogonal scene basis vectors are introduced as scene decoupling feature libraries. Each orthogonal scene basis vector consists of statistics including mean and variance, which are used to construct an orthogonal subspace of scene features. The similarity score is obtained by calculating the distribution distance between the scene features of the sample and the basic features in the scene decoupling feature library. The similarity score is converted into a matching weight, and multiple scene bases are weighted and fused to generate a mixed statistic for modulating the current sample. Adaptive instance normalization is used to align source domain image features to a mixture of statistics, thereby correcting for inter-domain differences.

2. The aircraft target domain adaptation detection method based on scene feature decoupling according to claim 1, characterized in that, The scene decoupling feature enhancer in the second step is used to continuously decouple and model image scene features during the image-text fusion process, specifically including: In each layer of the image-text interaction structure, the image features are first subjected to self-attention processing, and then bidirectional cross-attention calculation is performed with the text features. After each layer of image features is processed by the feedforward neural network, it is embedded in the scene decoupling module, and scene feature matching, fusion and modulation are repeatedly performed to ensure that the image features maintain decoupled representation in multi-layer semantic interaction; The text feature-guided decoupling mechanism enables the separability of scene features in different semantic directions of an image; The enhanced image-text fusion features after multi-layer decoupling are ultimately used to generate the detection query vector.

3. The aircraft target domain adaptation detection method based on scene feature decoupling according to claim 2, characterized in that, Bidirectional cross-attention includes image-to-text cross-attention and text-to-image cross-attention.

4. The aircraft target domain adaptation detection method based on scene feature decoupling according to claim 1, characterized in that, The dynamic expansion in the third step includes: The maximum matching score of the sample is calculated based on the similarity scores between the scene features of the current iteration sample and all existing scene bases. When the maximum matching score is lower than the set threshold, it is determined that the current sample cannot be fully modeled by the existing basis, triggering basis expansion, that is, registering the statistics of the sample as a new scene basis and adding it to the orthogonal subspace; The newly added basis participates in orthogonal constraint learning during training, and together with the original basis, they form the updated scene basis set.

5. The aircraft target domain adaptation detection method based on scene feature decoupling according to claim 1, characterized in that, The detection loss in the fourth step is the bounding box regression loss, which uses smooth L1 loss to measure the positional difference between the predicted box and the ground truth box, including the horizontal coordinate, vertical coordinate, width, height and rotation angle, and is calculated only for positive foreground samples; the contrast loss is the image-text contrast loss, which constrains each target query and the corresponding category text to be close to each other in the feature space, while being far away from other categories; the detection loss and contrast loss are jointly optimized to form a complete detection training objective function.

6. The aircraft target domain adaptation detection method based on scene feature decoupling according to claim 1, characterized in that, The enhanced domain image in the first step is constructed by performing brightness perturbation, color transformation, and contrast adjustment operations on the source domain image.

7. An aircraft target domain adaptation detection device based on scene feature decoupling, which implements the method of claim 1, characterized in that, include: The decoupling and modulation module takes the source domain image and its enhanced domain image as input, extracts multi-scale image features using shared weights, and embeds a scene decoupling image backbone network after each scale output. It modulates and aligns the feature representations of the source domain image and the enhanced domain image layer by layer to obtain image feature representations with domain adaptability. The semantic alignment and fusion module encodes the original text corresponding to the source domain image into text features through a text backbone network. The obtained image features and text features are then input into the scene decoupling feature enhancer, and deep fusion is performed through a multi-layer self-attention and cross-attention mechanism. The scene decoupling module is embedded in each branch after the deep fusion of image features to further enhance the domain invariance of image features and improve the semantic alignment capability of images and text. The scene enhancement module, in each scene decoupling feature enhancer, judges the coverage capability of the existing scene base set based on the matching similarity between the current sample scene features and the current scene base features; when a sample with a matching degree lower than the threshold is found, the statistic of the sample is registered as a new scene base direction, realizing the dynamic expansion of the scene feature space; The detection result generation module fuses image features and text features to perform language-guided query selection, and then inputs the results into a cross-modal decoder for target recognition and localization. During training, detection loss and contrast loss are used together for optimization, ultimately achieving cross-domain adaptive detection of aircraft targets.

8. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the aircraft target domain adaptation detection method based on scene feature decoupling as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to implement the aircraft target domain adaptation detection method based on scene feature decoupling as described in any one of claims 1 to 6.

Images