Remote sensing image plane target fine-grained classification method based on improved YOLO11
By improving the feature pyramid, edge sensing module, and classification head of the YOLO11 model, and combining them with a multi-task loss function, the accuracy and efficiency issues of aircraft target recognition in remote sensing images were resolved. This achieved fine-grained classification at the sub-aircraft type level, improving the recognition accuracy and real-time processing capabilities of aircraft targets in remote sensing images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2026-04-30
- Publication Date
- 2026-06-05
AI Technical Summary
Existing remote sensing image aircraft target recognition technologies suffer from problems such as loss of small target discrimination features, contradiction between inter-class similarity and intra-class differences, and interference from complex backgrounds in subdivided aircraft types. This results in low classification accuracy and high computational complexity, making it difficult to meet the needs of real-time remote sensing processing.
An improved YOLO11 model is adopted, which achieves fine-grained classification through feature pyramid construction, edge perception module, and classification head, combined with a multi-task loss function. The feature pyramid adopts a bidirectional cross-level attention fusion mechanism, the edge perception module extracts high-frequency components of the aircraft target contour, and the classification head dynamically generates category prototype vectors based on prototype contrastive learning. The model combines target localization, edge enhancement, and decoupled feature contrastive loss to coordinate multi-task learning.
It improves the accuracy and efficiency of fine-grained classification of aircraft targets in remote sensing images, solves the problems of weak features of small targets, background noise interference, and inter-class similarity and intra-class differences, and achieves sub-aircraft type level classification and recognition.
Abstract
Description
Technical Field
[0001] This invention relates to the field of fine-grained target recognition technology in remote sensing images, and in particular to a fine-grained classification method for aircraft targets in remote sensing images based on an improved YOLO11. Background Technology
[0002] Aircraft target recognition from remote sensing images has significant application value in the field of civil aviation management. Traditional methods rely on manually designed features (such as contour moments and texture statistics), and their classification accuracy is limited by changes in imaging conditions and the insufficient robustness of manually designed features. With the development of deep learning, algorithms based on single-stage detection frameworks have significantly improved the aircraft detection rate, but they still face three core shortcomings in subdivided aircraft types: First, there is the problem of lost features for small targets. Aircraft targets captured by remote sensing platforms occupy a very small proportion of the image. The downsampling mechanism of existing models causes key details (such as the number of engines and the shape of winglets) to be lost in the deep feature map, resulting in insufficient basis for subtype classification.
[0003] Secondly, there is a contradiction between inter-class similarity and intra-class differences. Different subtypes of aircraft are highly similar in overall structure, while aircraft of the same type exhibit diverse shapes due to differences in shooting angles. The coarse-grained features used by traditional detection models are insufficient to capture subtle discriminative features.
[0004] Third, interference from complex backgrounds. Airport scenarios contain similar-looking objects such as jet bridges and cargo vehicles. Existing methods fail to effectively suppress background responses during the feature extraction stage, leading to high-confidence false detections that interfere with fine-grained classification.
[0005] Current mainstream models improve classification performance by increasing network depth or introducing attention mechanisms, but this drastically increases computational complexity, making it difficult to meet the real-time processing requirements of remote sensing. Therefore, there is an urgent need to construct a lightweight, fine-grained classification architecture that balances accuracy and efficiency. Summary of the Invention
[0006] In view of this, the present invention proposes a fine-grained classification method for aircraft targets in remote sensing images based on an improved YOLO11, in order to solve the problems of low accuracy or low efficiency in fine-grained classification of aircraft in remote sensing images in the prior art.
[0007] The specific technical solution of this invention is as follows: A fine-grained classification method for aircraft targets in remote sensing images based on an improved YOLO11 includes: First, the remote sensing images are preprocessed, including image patch cropping, data augmentation, and pixel value normalization. Secondly, an improved YOLO11 model is used to process the preprocessed image. This model includes a feature pyramid construction module, an edge perception module, a feature decoding module, and a classification head. The feature pyramid construction module adopts a bidirectional cross-level attention fusion mechanism to adaptively weight and fuse deep semantic features with shallow detail features. The edge perception module extracts high-frequency components of the aircraft target contour and strengthens the edge features of key components by combining semantic edge constraint algorithms. The feature decoding module introduces a class-aware feature separation mechanism to dynamically separate the discriminative local features of different aircraft models from the global features. The classification head is based on prototype contrastive learning, dynamically generates category prototype vectors, and achieves fine-grained classification through feature similarity measurement and adaptive threshold decision-making. Finally, the target localization loss, edge enhancement loss, and decoupled feature contrast loss are combined, and a gradient balancing strategy is used to coordinate multi-task learning.
[0008] Specifically, the preprocessing operations include: cropping the original remote sensing image into fixed-size image blocks, retaining overlapping areas between adjacent image blocks; randomly flipping, rotating, adjusting brightness, and adjusting contrast of the cropped image blocks; and normalizing the pixel values of the image blocks, adjusting the pixel value range to between 0 and 1.
[0009] Specifically, the feature pyramid construction includes five feature levels with different resolutions, corresponding to the downsampling rates of the input image from shallow to deep. A bidirectional cross-level attention fusion mechanism is adopted, which includes two feature transfer paths: top-down and bottom-up. The top-down path upsamples deep features and concatenates them with shallow features of the same scale, while the bottom-up path downsamples shallow features and concatenates them with deep features of the same scale. The concatenated features are then input into the attention fusion module for adaptive weighting.
[0010] Specifically, the attention fusion module includes two parallel branches: a spatial attention branch and a channel attention branch. The spatial attention branch calculates the importance of each spatial location in the feature map to generate a spatial attention weight map, weighting the spatial dimension of the feature map to highlight the region where the aircraft target is located. The channel attention branch calculates the importance of each channel in the feature map to generate a channel attention weight vector, weighting the channel dimension of the feature map to highlight the channel features of the discriminative information of the aircraft target. The calculation process of both spatial attention weights and channel attention weights uses a normalized activation function. The weighted feature map is then output after the number of channels is adjusted by a convolutional layer.
[0011] Specifically, the edge perception module first extracts high-frequency components from the fused feature map and uses operator convolution to extract high-frequency edge information. The high-frequency component feature map is then concatenated with the original fused feature map and input into the semantic edge constraint algorithm. The semantic edge constraint algorithm classifies edge pixels into aircraft target edge pixels or background edge pixels based on the similarity between the semantic features around the edge pixel and the preset aircraft edge semantic features. The background edge pixel region is smoothed by mean filtering, and the aircraft target edge pixel region is enhanced by feature enhancement to improve the response intensity and highlight the edge features of key components.
[0012] Specifically, the class-aware feature separation mechanism includes: performing global average pooling on the optimized feature map to obtain a global feature vector; using a sliding window approach to divide the feature map into blocks, with each sliding window corresponding to a fixed-size local region, the window size being 1 / 16 of the feature map size and the step size being 1 / 2 of the window size; after transforming the features of each local region into local feature vectors, calculating the similarity with the global feature vector, and determining that the local feature vectors with similarity higher than a preset threshold are discriminative information and retained, while those with similarity lower than the threshold are filtered out; spatial channel decoupling operation performs spatial dimension encoding and channel dimension clustering on the local features to distinguish the features corresponding to different components, and finally splices the features of different components and fuses them with the global features for output.
[0013] Specifically, during the training phase, the classification head constructs a category prototype library, with each category corresponding to multiple prototype vectors, obtained by clustering discriminative local features of the training set samples. During the inference phase, discriminative local features of the input samples are extracted and similarity is measured with all prototype vectors in the prototype library to obtain the maximum similarity value for each category. The adaptive threshold decision dynamically adjusts the classification threshold based on the feature quality of the input samples. Feature quality is comprehensively evaluated by feature response strength, edge integrity, and local feature richness. If the maximum similarity value is higher than the classification threshold, it is determined to be a category; otherwise, it is determined to be an unknown category.
[0014] Specifically, the loss functions include target localization loss, edge enhancement loss, and decoupling feature contrast loss. The target localization loss uses the intersection-union loss function to optimize bounding box prediction; the edge enhancement loss uses the binary cross-entropy loss function to optimize edge extraction accuracy; and the decoupling feature contrast loss uses the contrast loss function to optimize feature separation and improve discriminative local feature similarity. The gradient balancing strategy dynamically adjusts the weights of the three loss functions, prioritizing target localization loss in the initial training stage, increasing the weight of edge enhancement loss in the middle stage, and increasing the weight of decoupling feature contrast loss in the later stage. After each training cycle, the weights are adjusted according to the validation set performance indicators to improve the task in a coordinated manner.
[0015] Specifically, the attention fusion module retains only the spatial attention branch or the channel attention branch; for remote sensing images with simple backgrounds, only the spatial attention branch is used; for remote sensing images with complex ground features, the channel attention branch is used to improve the accuracy of feature selection.
[0016] Specifically, the local region feature extraction adopts the region of interest alignment method instead of sliding window segmentation; the region of interest alignment is based on the aircraft target bounding box obtained by target detection to perform local region cropping, accurately locate the aircraft target region to avoid background interference.
[0017] The beneficial effects of this invention are as follows: 1. A multi-scale feature pyramid and a bidirectional cross-level attention fusion mechanism are used to enhance the feature response of small target aircraft by adaptively weighting and fusing deep semantic and shallow detail features, thus solving the problem that the features of small aircraft targets are weak and easily submerged by background noise.
[0018] 2. By using a high-frequency information-guided edge perception module and a semantic edge constraint algorithm, high-frequency components of the aircraft outline are extracted and edge features of key components are enhanced to suppress background noise, providing discriminative local feature basis for fine-grained classification.
[0019] 3. By using a class-aware feature separation mechanism and decoupling from the spatial channel, the discriminative local features of different aircraft models are dynamically separated, avoiding interference from redundant global feature information and improving the classification accuracy of similar aircraft models.
[0020] 4. By using a classification head based on prototype contrastive learning and adaptive threshold decision-making, class prototype vectors are dynamically generated. Through feature similarity measurement and adaptive threshold, fine-grained classification at the sub-machine level is achieved, solving the problem of inter-class similarity and intra-class differences.
[0021] 5. By coordinating multi-task learning through a gradient balancing strategy, and combining target localization, edge enhancement, and decoupled feature comparison losses, the loss weights are dynamically adjusted to improve the synergy between classification accuracy and localization robustness. Detailed Implementation
[0022] To make the technical problems, solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0023] This invention proposes a fine-grained classification method for aircraft targets in remote sensing images based on an improved YOLO11. This method is applied to intelligent interpretation scenarios of high-resolution remote sensing images and can achieve sub-aircraft type-level classification and recognition of aircraft parked in areas such as airports and ports.
[0024] The remote sensing images used in this invention come from the publicly available FAIR1M dataset. The dataset includes remote sensing images with a spatial resolution ranging from 0.3 meters to 0.8 meters, covering hundreds of typical airport areas worldwide. The labeled aircraft targets cover 17 fine-grained categories, including Boeing series, Airbus series, and domestic civil aircraft series. Each category contains at least 300 labeled samples, which can fully meet the needs of model training and testing.
[0025] Before model training, the remote sensing images are preprocessed, including: cropping the original remote sensing images into fixed-size image patches, retaining 15% overlap between adjacent image patches to avoid feature loss due to truncated aircraft targets; performing data augmentation operations such as random flipping, rotation, brightness adjustment, and contrast adjustment on the cropped image patches to expand the diversity of training samples and avoid overfitting problems in the model; and normalizing the pixel values of the image patches, uniformly adjusting the pixel value range to between 0 and 1 to improve the stability of model training.
[0026] The improved YOLO11 model used in this invention is based on the publicly available YOLO11 architecture. It retains the feature extraction capability of the backbone network in the original architecture, and at the same time, it makes targeted improvements to five parts, namely the feature pyramid, feature optimization module, feature decoding module, classification head and loss function, in order to achieve fine-grained classification.
[0027] The method of this invention first constructs a multi-scale feature pyramid for the input remote sensing image. Through a bidirectional cross-level attention fusion mechanism, it adaptively weights and fuses deep semantic features with shallow detail features to enhance the feature response of small target aircraft. This method is used to solve the problem that the features of small aircraft targets in remote sensing images are weak and easily submerged by background noise. Traditional feature pyramids usually only use a top-down feature transfer path, and deep semantic features cannot be effectively and fully fused with shallow detail features, resulting in insufficient feature response strength for small targets.
[0028] The multi-scale feature pyramid constructed in this step includes five feature levels with different resolutions, corresponding to downsampling rates of 1 / 8, 1 / 16, 1 / 32, 1 / 64, and 1 / 128 of the input image, from shallow to deep. The shallow features include more detailed information such as the contour and texture of the aircraft target, while the deep features include more global information such as the category and semantics of the aircraft target.
[0029] The bidirectional cross-level attention fusion mechanism includes two feature transfer paths: top-down and bottom-up. The top-down path upsamples deep features and concatenates them with shallow features of the same scale, while the bottom-up path downsamples shallow features and concatenates them with deep features of the same scale. The output features of the two paths are concatenated and then input into the attention fusion module for adaptive weighting.
[0030] The attention fusion module comprises two parallel branches: a spatial attention branch and a channel attention branch. The spatial attention branch calculates the importance of each spatial location in the feature map, generating a spatial attention weight map. This weights the spatial dimensions of the feature map, highlighting the feature responses of the region containing the aircraft target and suppressing features from the background region. The channel attention branch calculates the importance of each channel in the feature map, generating a channel attention weight vector. This weights the channel dimensions of the feature map, highlighting channel features containing discriminative information about the aircraft target and suppressing interference from redundant channels. Both spatial and channel attention weights are normalized using the sigmoid activation function, ensuring the weight values range from 0 to 1. The weighted feature map is then processed through convolutional layers to adjust the number of channels before being output to subsequent feature optimization steps.
[0031] As another embodiment of the present invention, the attention fusion module may retain only the spatial attention branch or the channel attention branch. For remote sensing images of airport areas with relatively simple backgrounds, the spatial attention branch alone can meet the requirements of feature enhancement, further simplifying the model structure. For remote sensing images with complex ground features, the channel attention branch can be superimposed to improve the accuracy of feature selection.
[0032] Next, the high-frequency information-guided edge perception module is used to extract the high-frequency components of the aircraft target outline. Combined with the semantic edge constraint algorithm, background noise is suppressed and the edge features of key components such as wings and tail are enhanced. The purpose is to enhance the edge features of key components of the aircraft target, so as to provide discriminative local feature basis for subsequent fine-grained classification. Traditional feature extraction methods usually process edge features and global features in a unified manner, which makes the edge features of key components easy to be weakened and unable to distinguish different aircraft models with similar shapes.
[0033] The high-frequency information-guided edge perception module in this invention first extracts high-frequency components from the input fused feature map. Specifically, it uses the Laplacian operator to perform convolution operations on the feature map to extract high-frequency edge information from the feature map. The high-frequency components mainly correspond to the features of the aircraft target outline, wing edge, tail edge, engine outline, and other regions.
[0034] After the extracted high-frequency component feature map is concatenated with the original fused feature map, it is input into the semantic edge constraint algorithm for optimization. The semantic edge constraint algorithm first classifies the edge pixels in the feature map into aircraft target edge pixels and background edge pixels. The classification is based on the similarity between the semantic features around the edge pixel and the pre-trained semantic features of the aircraft edge. Pixels with a similarity higher than a preset threshold are determined to be aircraft target edge pixels, and pixels with a similarity lower than the preset threshold are determined to be background edge pixels. For regions determined to be background edge pixels, mean filtering is used for feature smoothing to suppress background noise interference. For regions determined to be aircraft target edge pixels, feature enhancement is used to increase the feature response intensity of the region by 1.5 times, highlighting the edge features of key components such as wings, tail fins, and engines. For large aircraft targets with obvious edge features, the enhancement factor of the edge features can be appropriately reduced to avoid feature overfitting. For small aircraft targets with weak edge features, the enhancement factor of the edge features can be appropriately increased to ensure that the response intensity of the edge features meets the requirements of subsequent classification.
[0035] Among them, the semantic edge constraint algorithm can use a pre-trained edge detection model to classify edge pixels. The edge detection model is trained with a small number of labeled edge samples, which can further improve the accuracy of edge pixel classification and reduce the probability of misjudgment. It is especially suitable for remote sensing images of complex scenes containing a large number of linear ground objects such as buildings and roads.
[0036] Subsequently, a class-aware feature separation mechanism is introduced in the feature decoding stage. Through spatial channel decoupling operation, the discriminative local features of different aircraft models (such as engine layout and landing gear structure) are dynamically separated from the global features. This is to avoid redundant information in the global features from interfering with fine-grained classification. Traditional feature decoding methods usually classify the global features directly, which cannot make full use of the discriminative information of local key components, resulting in low classification accuracy of similar aircraft models.
[0037] The class-aware feature separation mechanism first performs global average pooling on the optimized input feature map to obtain a global feature vector, which includes the overall category information of the aircraft target. Simultaneously, it extracts local region features from the feature map using a sliding window approach, dividing the feature map into blocks. Each sliding window corresponds to a fixed-size local region in the original remote sensing image. In this embodiment, the size of the sliding window is set to 1 / 16 of the feature map size, and the stride is set to 1 / 2 of the sliding window size, ensuring sufficient overlap between each local region and preventing the omission of key component features. The features of each local region are transformed by a convolutional layer to obtain a local feature vector. The similarity between all local feature vectors and the global feature vector is calculated using cosine similarity. Local feature vectors with similarity scores higher than a preset threshold are considered to contain discriminative information, while those lower than the preset threshold are considered redundant features and filtered out.
[0038] As another embodiment, local region feature extraction can replace sliding window segmentation with region of interest alignment. Region of interest alignment is based on the aircraft target bounding box obtained by target detection to perform local region cropping, which can accurately locate the region where the aircraft target is located, avoid interference from background region features, and further improve the purity of local features.
[0039] The spatial channel decoupling operation first decouples the selected local features spatially, encodes the spatial position of each local feature, records the corresponding region position in the original image, and distinguishes the regions where different components such as wings, tail, fuselage, and engines are located. Then, channel dimension decoupling is performed, and the channel information of each local feature is clustered. Channels corresponding to the same type of component features are grouped to obtain feature channel groups corresponding to different components.
[0040] Finally, the features corresponding to different components are concatenated to obtain the separated discriminative local features, which are then fused with the global features and output to the subsequent classification steps.
[0041] Next, a classification head based on prototype contrastive learning is constructed to dynamically generate category prototype vectors. Through feature similarity measurement and adaptive threshold decision-making, fine-grained classification at the sub-aircraft type level is achieved, while simultaneously addressing the challenges of inter-class similarity and intra-class differences. The core objective of this step is to improve the classification accuracy of similar aircraft models. Traditional classification heads typically use fixed classification weights for prediction, which cannot adapt to intra-class differences of the same aircraft model under different shooting angles and imaging conditions, as well as inter-class similarity issues caused by the similar shapes of different aircraft models.
[0042] The prototype-based contrastive learning classifier first constructs a category prototype library during the training phase. Each category corresponds to multiple prototype vectors. The prototype vectors are obtained by clustering the discriminative local features of all samples of that category in the training set. The clustering uses the k-means clustering algorithm. Each category is set with 3 prototype vectors, which correspond to the feature distribution of the aircraft of that category from three different shooting angles: front, side, and top.
[0043] During the inference phase, the discriminative local features of the input sample are first extracted, and then their similarity is measured with all prototype vectors in the category prototype library. The similarity measurement adopts the cosine similarity method to obtain the maximum similarity value for each category.
[0044] The adaptive threshold decision-making process first dynamically adjusts the classification threshold based on the feature quality of the input samples. Feature quality is comprehensively evaluated using three indicators: feature response strength, edge integrity, and local feature richness. The higher the evaluation score, the lower the classification threshold is set; conversely, the lower the evaluation score, the higher the classification threshold is set. If the maximum similarity value of a certain category is higher than the classification threshold, the input sample is determined to belong to that category; if the maximum similarity value of all categories is lower than the classification threshold, it is determined to be an unknown category, thus avoiding misclassification.
[0045] Finally, by combining the target localization loss, edge enhancement loss, and decoupled feature contrast loss, a gradient balancing strategy is employed to coordinate multi-task learning, improving the synergy between classification accuracy and localization robustness. This step aims to coordinate the optimization directions of multiple tasks, preventing over-optimization of a single task from degrading the performance of other tasks. Traditional multi-task learning typically uses fixed loss weights, which cannot adapt to the task optimization needs of different training stages, resulting in the model's overall performance failing to reach its optimal level.
[0046] The target localization loss uses the intersection-union loss function to optimize the prediction accuracy of the aircraft target bounding box and ensure accurate localization. The edge enhancement loss uses the binary cross-entropy loss function to optimize the edge extraction accuracy of the edge perception module and improve the response intensity of edge features of key components. The decoupling feature contrast loss uses the contrast loss function to optimize the feature separation effect of the fine-grained feature decoupling step, making the similarity between discriminative local features of the same category higher and the similarity between discriminative local features of different categories lower, thereby improving the discriminativeness of the features.
[0047] The gradient balancing strategy dynamically adjusts the weights of the three loss functions during training. Specifically, in the initial training phase, the weight of the target localization loss is set to the highest to ensure the model can accurately detect the aircraft target's location. In the middle training phase, the weight of the edge enhancement loss is increased to optimize edge feature extraction. In the later training phase, the weight of the decoupling feature contrast loss is increased to optimize feature discriminativeness and improve the accuracy of fine-grained classification. After each training epoch, the validation set performance metrics for the three tasks are calculated. If the performance metric of a certain task stops improving, the loss weight of that task is appropriately reduced, while the loss weights of the other tasks are increased to ensure coordinated performance improvement across all tasks.
[0048] The beneficial effects of this invention are as follows: 1. By constructing a multi-scale feature pyramid and combining it with a bidirectional cross-level attention fusion mechanism, deep semantic features and shallow detail features are adaptively weighted and fused, significantly enhancing the feature response intensity of small-sized aircraft targets. This mechanism effectively solves the problem that the features of small targets are weak and easily submerged by background noise. Compared with the traditional unidirectional path, the bidirectional path design improves the integrity of feature fusion and ensures the stability of small-target aircraft recognition in complex backgrounds.
[0049] 2. A high-frequency information-guided edge perception module extracts the high-frequency components of the aircraft outline. Combined with a semantic edge constraint algorithm, this enhances the edge features of key components such as wings and tail fins while suppressing background noise interference. This effect provides discriminative local feature basis for fine-grained classification. By distinguishing the aircraft target edge from the background edge, it improves the saliency of key component features and avoids edge weakening problems in global feature processing.
[0050] 3. The class-aware feature separation mechanism dynamically separates discriminative local features of different aircraft models, such as engine layout and landing gear structure, through spatial channel decoupling, effectively avoiding interference from redundant information in global features. This effect improves the classification accuracy of similar aircraft models, and enhances the model's ability to distinguish subtle differences through local feature extraction and similarity filtering.
[0051] 4. A classification head based on prototype contrastive learning is constructed to dynamically generate category prototype vectors. Through feature similarity measurement and adaptive threshold decision-making, fine-grained classification at the sub-camera model level is achieved. This effect solves the challenges of inter-class similarity and intra-class differences. By constructing a prototype library and adjusting the threshold, it adapts to different shooting angles and imaging conditions, improving classification accuracy and robustness.
[0052] 5. A gradient balancing strategy combines target localization loss, edge enhancement loss, and decoupled feature contrast loss to dynamically adjust loss weights and coordinate the multi-task learning process. This improves the synergy between classification accuracy and localization robustness. Through weight optimization during the training phase, it ensures balanced performance improvement across tasks, avoiding overall performance degradation caused by a single task dominating.
[0053] The above are merely preferred embodiments of the present invention and are 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 protection scope of the present invention.
Claims
1. A fine-grained classification method for aircraft targets in remote sensing images based on an improved YOLO11, characterized in that, include: First, the remote sensing images are preprocessed, including image patch cropping, data augmentation, and pixel value normalization. Secondly, the preprocessed image is processed using an improved YOLO11 model, which includes a feature pyramid construction, an edge perception module, a feature decoding module, and a classification head. The feature pyramid construction adopts a bidirectional cross-level attention fusion mechanism to adaptively weight and fuse deep semantic features with shallow detail features. The edge perception module extracts the high-frequency components of the aircraft target outline and combines them with a semantic edge constraint algorithm to enhance the edge features of key components. The feature decoding module introduces a class-aware feature separation mechanism to dynamically separate the discriminative local features of different aircraft models from the global features; the classification head is based on prototype contrastive learning to dynamically generate category prototype vectors and achieve fine-grained classification through feature similarity measurement and adaptive threshold decision-making. Finally, the target localization loss, edge enhancement loss, and decoupled feature contrast loss are combined, and a gradient balancing strategy is used to coordinate multi-task learning.
2. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The preprocessing operations specifically include: cropping the original remote sensing image into image blocks of fixed size, retaining overlapping areas between adjacent image blocks; randomly flipping, rotating, adjusting the brightness and contrast of the cropped image blocks; and normalizing the pixel values of the image blocks, adjusting the pixel value range to between 0 and 1.
3. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The feature pyramid construction includes five feature levels with different resolutions, corresponding to the downsampling rates of the input image from shallow to deep. A bidirectional cross-level attention fusion mechanism is adopted, which includes two feature transfer paths: top-down and bottom-up. The top-down path upsamples deep features and concatenates them with shallow features of the same scale, while the bottom-up path downsamples shallow features and concatenates them with deep features of the same scale. The concatenated features are then input into the attention fusion module for adaptive weighting.
4. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The attention fusion module includes two parallel branches: a spatial attention branch and a channel attention branch. The spatial attention branch calculates the importance of each spatial location in the feature map to generate a spatial attention weight map, and weights the spatial dimensions of the feature map to highlight the region where the aircraft target is located. The channel attention branch calculates the importance of each channel in the feature map to generate a channel attention weight vector, and weights the channel dimensions of the feature map to highlight the channel features of the aircraft target discriminative information. Both spatial attention weights and channel attention weights are calculated using normalized activation functions. The weighted feature maps are then processed by convolutional layers to adjust the number of channels before being output.
5. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The edge perception module first extracts high-frequency components from the fused feature map and uses operator convolution to extract high-frequency edge information. The high-frequency component feature map is then concatenated with the original fused feature map and input into the semantic edge constraint algorithm. The semantic edge constraint algorithm classifies edge pixels as aircraft target edge pixels or background edge pixels based on the similarity between the semantic features around the edge pixel and the preset aircraft edge semantic features. Mean filtering is applied to smooth the background edge pixel region, and feature enhancement processing is performed on the aircraft target edge pixel region to improve the response intensity and highlight the edge features of key components.
6. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The class-aware feature separation mechanism includes: performing global average pooling on the optimized feature map to obtain a global feature vector; using a sliding window approach to divide the feature map into blocks, with each sliding window corresponding to a fixed-size local region, the window size being 1 / 16 of the feature map size and the step size being 1 / 2 of the window size; after transforming the features of each local region into local feature vectors, calculating the similarity with the global feature vector, and determining that the local feature vectors with similarity higher than a preset threshold are discriminative information and retained, while those with similarity lower than the threshold are filtered out; spatial channel decoupling operation performs spatial dimension encoding and channel dimension clustering on the local features to distinguish the features corresponding to different components, and finally splicing the features of different components and fusing them with the global features for output.
7. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The classification head constructs a category prototype library during the training phase, with each category corresponding to multiple prototype vectors, which are obtained by clustering discriminative local features of the training set samples. During the inference phase, discriminative local features of the input samples are extracted and similarity measures are performed with all prototype vectors in the prototype library to obtain the maximum similarity value for each category. The adaptive threshold decision dynamically adjusts the classification threshold based on the feature quality of the input samples. Feature quality is comprehensively evaluated by feature response intensity, edge integrity, and local feature richness. If the maximum similarity value is higher than the classification threshold, it is determined to be of that category; otherwise, it is determined to be of an unknown category.
8. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The loss function includes target localization loss, edge enhancement loss, and decoupling feature contrast loss. The target localization loss uses the intersection-union loss function to optimize bounding box prediction; the edge enhancement loss uses the binary cross-entropy loss function to optimize edge extraction accuracy; and the decoupling feature contrast loss uses the contrast loss function to optimize feature separation and improve discriminative local feature similarity. A gradient balancing strategy dynamically adjusts the weights of the three loss functions. In the initial stage of training, the target localization loss is prioritized, the edge enhancement loss weight is increased in the middle stage, and the decoupling feature contrast loss weight is increased in the later stage. After each training cycle, the weights are adjusted according to the validation set performance indicators to improve the task in a coordinated manner.
9. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 1, characterized in that, The attention fusion module retains only the spatial attention branch or the channel attention branch; For remote sensing images with simple backgrounds, only the spatial attention branch is used; for remote sensing images with complex ground features and interference, the channel attention branch is used to improve the accuracy of feature selection.
10. The fine-grained classification method for aircraft targets in remote sensing images as described in claim 5, characterized in that, The local region feature extraction adopts the region of interest alignment method to replace the sliding window block; the region of interest alignment is based on the aircraft target bounding box obtained by target detection to perform local region cropping, accurately locate the aircraft target region to avoid background interference.