Remote sensing image aircraft detection method and system based on enhanced fine-grained features

Abstract

The invention discloses a remote sensing image aircraft detection method and system based on enhanced fine-grained features, and relates to the technical field of image processing. The objective of the invention is to solve the problems that features are easy to lose and background interference is strong when small, dense and multi-scale aircraft targets are detected in the prior art. The core of the method is that a WS-YOLO detection network is constructed, and an innovative enhanced fine-grained module is embedded in a neck network of the WS-YOLO detection network; the module performs progressive fine-grained feature extraction and channel importance re-calibration in parallel through a double-branch design, and fuses multi-granularity information, so that key detail features of an aircraft are effectively reserved and enhanced, and meanwhile, complex background interference is inhibited; according to the method, multiple loss functions are fused in training to optimize the model, and on the premise that the efficient reasoning speed of a single-stage detector is kept, the detection precision and scene robustness of aircrafts in remote sensing images, especially small-size targets, are remarkably improved, and the method has good practical value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a remote sensing image aircraft detection method and system based on enhanced fine-grained features. BACKGROUND

[0002] Satellite remote sensing image-based aircraft detection technology has gradually become a research hotspot and is widely used in resource monitoring, urban planning, environmental protection, natural disaster warning and other fields. Aircraft as a high-value and relatively small target, its accurate identification is of great significance in target detection field.

[0003] Traditional aircraft detection methods rely heavily on complex feature representation techniques or parameter optimization made by artificial methods. Aircraft targets in remote sensing images are usually small in size, weak in features, and often densely distributed in complex scenes such as airports and hangars, which have serious background interference, scale variation and large intra-class differences. This makes it difficult for traditional artificial feature-based methods and general deep learning detectors to achieve high-precision and high-robust detection.

[0004] In the existing mainstream technology, single-stage detectors based on YOLO series are applied to this task due to their fast speed, but their default structure has limitations in processing fine-grained features of remote sensing images. Specifically: 1) In the multiple downsampling process, the fine features of the aircraft are easy to be lost; 2) The general feature fusion network is difficult to effectively suppress background noise similar to the texture of the aircraft (such as building roofs and ground vehicles); 3) Lack of special structure design for small target high-precision positioning and identification.

[0005] Therefore, there is an urgent need for a special aircraft detection scheme that can enhance fine-grained feature representation, suppress background interference, and maintain high inference speed to solve the problems existing in the prior art. SUMMARY

[0006] The purpose of the present application is to overcome the shortcomings of the prior art and provide a remote sensing image aircraft detection method and system based on enhanced fine-grained features, aiming to propose a special detection architecture of WS-YOLO, the innovation point of which is concentrated in the proposed "enhanced fine-grained module (EFM)". This module is embedded in the neck of the detection network to refine and enhance the fine spatial structure and channel features related to the aircraft from multi-scale features.

[0007] To achieve the above purpose, the present application provides a remote sensing image aircraft detection method based on enhanced fine-grained features, comprising the following steps: Step S1: Construct the WS-YOLO detection network; wherein the WS-YOLO detection network is a single-stage detector, comprising a backbone network, a neck network, and a detection head; the backbone network is divided into four extraction stages; Step S2: Construct an enhanced fine-grained module; wherein, the extraction stage uses convolutional layers, BN layers, and ReLU layers, and the enhanced fine-grained module is added after the ReLU layers of each extraction stage of the backbone network; Step S3: Use the remote sensing image dataset containing aircraft annotations to train the WS-YOLO detection network end-to-end to obtain the trained aircraft detection model; Step S4: Input the remote sensing image to be detected into the trained aircraft detection model, and output the bounding box position, category label and confidence score of the aircraft in the image; The processing flow of the enhanced fine-grained module includes: The input feature map is equally divided into first branch features along the channel dimension. Second branch features ; For the first branch feature Progressive fine-grained extraction is performed to obtain several intermediate features of the first branch; For the second branch features Perform channel importance recalibration to obtain the second branch calibration features; All intermediate features from the first branch and the calibrated features from the second branch are concatenated along the channel dimension to generate an enhanced fine-grained feature map.

[0008] As a further solution, the backbone network learns discriminative aircraft features and processes them sequentially through four extraction stages of downsampling and fine-grained enhancement modules, obtaining feature maps S1, S2, S3, and S4 in hierarchical order.

[0009] As a further solution, a context enhancement layer is also set up to process the deepest feature map S4 output by the backbone network. The feature map S4 is then processed by passing it through the SPPF module and C2PSA module in the YOLO model to generate high-level features Fs rich in global context information.

[0010] As a further solution, the neck network is used to perform multi-scale feature fusion and fine-grained enhancement, the specific process of which includes: The high-level features Fs, feature map S2, and feature map S3 are input into the neck network; The high-level feature Fs is continuously upsampled to obtain upsampled features at different scales; Multi-scale feature fusion is performed based on upsampled features at different scales, feature map S2 and feature map S3; Based on the fusion of aircraft features at different scales, cross-level information fusion features N1, N2, N3 and N4 are obtained.

[0011] As a further solution, the detection head is provided with three detection head layers D1, D2 and D3 of different sizes corresponding to the three target scales of large, medium and small; wherein the fusion feature N2, fusion feature N3 and fusion feature N4 are respectively input to the detection head layer D3, detection head layer D2 and detection head layer D1.

[0012] As a further solution, the neck network inputs multi-scale fused features into the detection head, specifically including: The high-level feature Fs is upsampled so that its spatial size is consistent with the feature map S3 and then fused to obtain the cross-level information fusion feature N1. The cross-level information fusion feature N1 is upsampled so that its spatial size is consistent with the feature map S2 and then fused to obtain the fused feature N2. The fusion feature N2 is upsampled so that its spatial size is consistent with the cross-level information fusion feature N1, and then fused to obtain the fusion feature N3. The high-level feature Fs and the fusion feature N3 are upsampled and fused to obtain the fusion feature N4; The fusion features N2, N3, and N4 are respectively input into the detection head layer D1, detection head layer D2, and detection head layer D3.

[0013] As a further solution, for the first branch feature The process of progressive fine-grained extraction includes: The first branch feature The input is fed into the RepConv layer to obtain the initial fine-grained features. ; The initial fine-grained features Input a densely connected block containing n-1 intermediate convolutional layers, and perform deep feature extraction to obtain n-1 intermediate features of the first branch; The last intermediate feature from the first branch is channel-fused using a 1×1 convolutional layer to obtain the final feature map. .

[0014] As a further solution, the second branch feature The channel importance recalibration process includes: For the second branch features Adaptive average pooling and adaptive max pooling are performed separately to aggregate spatial information; After adding the two pooling results, input them into a channel attention generation network consisting of a multilayer perceptron and a sigmoid activation function to generate a channel attention weight vector. The channel attention weight vector is multiplied channel by channel with the original second branch features to obtain the recalibrated features.

[0015] As a further solution, the training process in step S3 uses a loss function that is a weighted sum of bounding box regression loss, classification loss, and distribution focal loss; wherein the bounding box regression loss is Wise-IoU Loss, the classification loss is BCE Loss, and the distribution focal loss is Distribution Focal Loss.

[0016] On the other hand, the present invention also provides a remote sensing image aircraft detection system based on enhanced fine-grained features, characterized in that it includes: The preprocessing module is used to receive and preprocess the input remote sensing images; The detection module integrates an aircraft detection network constructed and trained according to a remote sensing image aircraft detection method based on enhanced fine-grained features as described above, for extracting features from the preprocessed image and detecting aircraft targets. The post-processing and output module is used to perform non-maximum suppression processing on the raw output of the detection module and generate the final detection result visualization information.

[0017] Compared with related technologies, the aircraft detection method and system based on enhanced fine-grained features in remote sensing images provided by this invention have the following advantages: 1. The core Enhanced Fine-Grained Module (EFM) of this invention works in tandem with two branches: "progressive fine-grained extraction" and "channel importance recalibration". It can effectively preserve and enhance the fine structure and texture features of aircraft from multi-scale features. This invention can effectively alleviate the problem of small target feature loss caused by downsampling and significantly improve the detection capability of small and dense aircraft targets.

[0018] 2. The channel attention mechanism in the EFM module of this invention can adaptively recalibrate the weights of feature channels, suppress background noise channels that are irrelevant to aircraft, and enhance the expression of key feature channels. This enables the model to exhibit more stable detection performance and stronger generalization ability when facing complex airport scenes, different lighting conditions and sensor differences.

[0019] 3. The present invention embeds an EFM module in the neck network, which adopts a lightweight dual-branch design and introduces limited additional parameters and computational overhead; this efficient design enables the present invention to meet the application requirements of real-time or near-real-time processing of remote sensing images.

[0020] 4. The EFM module proposed in this invention has a clear structure and can be used as a plug-and-play enhancement unit. It can be flexibly embedded into the neck or backbone of existing single-stage detection networks (such as the YOLO series) without changing the overall network framework, thereby improving the detection performance of fine-grained targets.

[0021] In summary, this invention, through its innovative EFM module, fundamentally improves the feature extraction quality and discrimination capability of aircraft targets in remote sensing images without sacrificing inference efficiency. It achieves high-precision, high-robust, and high-efficiency aircraft detection, demonstrating good practical value and promising prospects for widespread application. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A schematic diagram illustrating the steps of a remote sensing image aircraft detection method based on enhanced fine-grained features provided by the present invention; Figure 2 This is a schematic diagram of the overall network architecture of WS-YOLO provided by the present invention; Figure 3 A detailed structural block diagram of the enhanced fine-grained module provided by the present invention; Figure 4 A schematic diagram of the structure of a remote sensing image aircraft detection system based on enhanced fine-grained features provided by the present invention; Figure 5 This is a diagram showing the detection results of this invention on the CORS-ADD dataset; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Please see Figure 1 This embodiment provides a remote sensing image aircraft detection method based on enhanced fine-grained features, including the following steps: Step S1: Construct the WS-YOLO detection network; wherein the WS-YOLO detection network is a single-stage detector, comprising a backbone network, a neck network, and a detection head; the backbone network is divided into four extraction stages; Step S2: Construct an enhanced fine-grained module; wherein, the extraction stage uses convolutional layers, BN layers, and ReLU layers, and the enhanced fine-grained module is added after the ReLU layers of each extraction stage of the backbone network; Step S3: Use the remote sensing image dataset containing aircraft annotations to train the WS-YOLO detection network end-to-end to obtain the trained aircraft detection model; Step S4: Input the remote sensing image to be detected into the trained aircraft detection model, and output the bounding box position, category label and confidence score of the aircraft in the image; The processing flow of the enhanced fine-grained module includes: The input feature map is equally divided into first branch features along the channel dimension. Second branch features ; For the first branch feature Progressive fine-grained extraction is performed to obtain several intermediate features of the first branch; For the second branch features Perform channel importance recalibration to obtain the second branch calibration features; All intermediate features from the first branch and the calibrated features from the second branch are concatenated along the channel dimension to generate an enhanced fine-grained feature map.

[0027] It should be noted that this embodiment benefits from the Enhanced Fine-grained Module (EFM), which effectively preserves and enhances the fine-grained features of aircraft and suppresses complex backgrounds, thereby improving the model's feature extraction capability for small aircraft targets. The EFM employs a dual-branch and lightweight design, significantly improving the model's detection accuracy with limited additional parameters and computational cost. The channel attention mechanism within the module enables the model to adaptively focus on key feature channels, improving detection stability under different sensors, lighting conditions, and background complexities. The model maintains the high-speed advantage of a single-stage detector, meeting the needs of real-time or near-real-time processing of remote sensing images. It is suitable for aircraft detection tasks in high-resolution remote sensing images, demonstrating good practicality and potential for wider application.

[0028] In step S1, the WS-YOLO detection network is a single-stage detector, which includes, in sequence: Backbone network: Learning distinctive aircraft features Where H×W represents the spatial dimension of the feature map and C represents the number of channels, it is used to extract multi-level features from the input remote sensing image. The input image is processed sequentially through four stages of downsampling and enhancement fine-grained module (EFM) to obtain feature maps S1, S2, S3, and S4.

[0029] Context enhancement layer: Processes the deepest feature map S4 output by the backbone network, and then passes it through the SPPF module and C2PSA module in the YOLO model to generate high-level features Fs rich in global context information.

[0030] Neck Network: Used for multi-scale feature fusion and fine-grained enhancement. Feature maps Fs, S2, and S3 are fed into the neck network for multi-scale fusion. High-level features Fs are successively upsampled N1, N2, N3, and N4 to restore them to the same dimensions as the specified input. Skip connections are applied between S2 and N2, S3 and N1, Fs and N4, and N1 and N3 to better fuse multi-scale aircraft features.

[0031] Detection head: Receives multi-scale feature maps after EFM processing and predicts the bounding boxes, categories, and confidence levels of aircraft at different scales in parallel.

[0032] Furthermore, the specific structural design and workflow of the Enhanced Fine-Grained Module (EFM) are key innovations of this invention, and its processing flow is as follows: 1) Two-branch feature separation: separating input features It is equally divided into two branches, Y1 and Y2, along the channel dimension: First Branch Second branch This design reduces the number of parameters and computational complexity of subsequent parallel processing.

[0033] 2) Progressive fine-grained extraction: for the first branch Progressive feature refinement is performed. First, it is processed through a RepConv layer. To obtain the initial fine-grained features The RepConv layer typically consists of a 3x3 convolution, a 1x1 convolution, and an identity connection combined into a single convolutional layer. Subsequently, it is passed through a layer containing... Deep feature extraction is performed using densely connected blocks of convolutional layers.

[0034] This represents the output of the k-th intermediate convolutional layer (out of a total of n-1 layers). Layer output From the output of the previous layer It is obtained through 3×3 convolution operation to gradually extract deeper features while mitigating the loss of features of small targets.

[0035] Finally, a 1×1 convolutional layer is used to perform channel fusion on the output to obtain the final output of this branch. This represents the final feature map of this branch, which enhances the expressive power of the features by mixing information across channels.

[0036] 3) Channel importance recalibration: for the second branch A channel attention mechanism is applied to adaptively emphasize information-rich channel features and suppress irrelevant or noisy channels. The specific process is as follows: [The text abruptly ends here, so the translation stops as well.] Adaptive average pooling (Avg) and adaptive max pooling (Max) are performed separately to aggregate spatial information. The results of the two pooling operations are then fed into a multilayer perceptron (MLP) consisting of two convolutional layers and the SiLU function, and finally activated by the Sigmoid activation function. Generate channel attention weight vectors. Then, combine these weight vectors with the original features. By performing channel-by-channel multiplication, the enhanced features are obtained: This represents element-wise multiplication. This represents the Sigmoid activation function, and Avg and Max represent average pooling and max pooling, respectively.

[0037] 4) Multi-granularity feature fusion: This involves fusing all intermediate features extracted from the first branch ( Features after calibration with the second branch Concatenate along the channel dimension to form the final enhanced fine-grained feature map: This indicates feature layer fusion and splicing. Z represents the output of the k-th intermediate convolutional layer (out of n - 1 layers), and Z represents the feature map extracted by the enhanced fine-grained module. This fusion strategy aggregates fine-grained information from different levels from low to high, as well as channel recalibration information.

[0038] Furthermore, during model training, the WS-YOLO network was trained end-to-end using a remote sensing image dataset containing aircraft annotations. The loss function consisted of a weighted average of bounding box regression loss (Wise-IoU Loss), classification loss (BCELoss), and distribution focal loss (Distribution Focal Loss). A stochastic gradient descent optimizer was used for optimization.

[0039] The bounding box regression loss function is used to precisely adjust the predicted box to be close to the true box.

[0040] Note: γ is the scaling factor, and β controls the decay rate.

[0041] The classification loss function is used to evaluate the quality of the prediction results of a binary classification model.

[0042] Note: y represents the true label {0, 1}, p represents the positive class probability predicted by the model, which is usually obtained through the Sigmoid activation function, and log is the natural logarithm to the base e.

[0043] The distribution-focusing loss function is a loss function used for object detection that addresses the imbalance between positive and negative samples by incorporating distribution information.

[0044] Note: z represents the vector of the neural network for a certain target, j represents the index of a certain position in the discrete interval, and λ represents the interpolation weight of the true value between two adjacent discrete points.

[0045] Finally, the remote sensing image to be detected is input into the trained WS-YOLO model, and the network automatically outputs the bounding box positions, class labels, and confidence scores of all aircraft in the image.

[0046] In one specific embodiment, a training input image of size 640×640 is used as an example.

[0047] Network construction example: The overall network structure of the WS-YOLO of this invention is mainly divided into three parts: the backbone network, the neck network, and the detection head layer.

[0048] 1) such as Figure 2 As shown, the backbone network uses the YOLO backbone to extract features.

[0049] It consists of four extraction stages, each using a 3×3 convolutional layer, a BN layer, and a ReLU layer. An Enhanced Fine-Grained Module (EFM) is added after the ReLU layer to output feature maps at four levels: S1 (128×160×160), S2 (256×80×80), S3 (512×40×40), and S4 (1024×20×20).

[0050] like Figure 3 As shown, the Enhanced Fine-Grained Module (EFM) consists of three branches: Progressive Fine-Grained Extraction, Channel Importance Recalibration, and a 1×1 convolutional network convergence acceleration layer. The features input to EFM are equally split into Y1 and Y2 along the channel dimension and fed into the Progressive Fine-Grained Extraction branch and the Channel Importance Recalibration branch.

[0051] The progressive fine-grained extraction branch Y1 first passes through the RepConv layer to obtain Z0, and then passes through a cascaded 3×3 convolution to obtain Zn. It is worth noting that the features Z0, Z1, Z3, etc. output by the RepConv layer and the cascaded convolution will be concatenated with the last layer feature Zn through channel splicing in the form of residuals.

[0052] The channel importance recalibration branch Y2 dynamically assigns weights to different feature positions through adaptive average pooling and adaptive max pooling. The results of average and max pooling are then subjected to double convolution and SiLU linear transformation, and channel weights are generated through the Sigmoid activation function. Finally, the channel weights are multiplied with the original feature Y2 channel by channel to obtain the feature Y2att.

[0053] Finally, a 1×1 accelerated convergence layer is used to concatenate the features extracted by progressive fine-grained extraction and the features after channel importance recalibration, aggregating the entire enhanced fine-grained module branch and outputting the enhanced fine-grained feature Z.

[0054] 2) Apply SPPF and C2PSA modules sequentially to the highest-dimensional feature map S4 (1024×20×20) to enhance the attention to aircraft features through the attention mechanism, and obtain feature Fs (1024×20×20).

[0055] 3) The neck network receives S2, S3, and Fs, and performs upsampling and fusion stages sequentially based on Fs.

[0056] First, Fs is upsampled to the same size as S3 to obtain the cross-level information fusion feature N1. Then, N1 and S2 are upsampled to the same size to obtain the fusion feature N2. Here, the fusion feature N2 is output as the detection head D3.

[0057] Then, N2 is upsampled to the same size as N1 and fused to obtain N3. The fused feature N3 is output as the detection head D2. Then, N4 and Fs are upsampled and fused to obtain N4. The fused feature N4 is output as the detection head D1.

[0058] 4) Output N2, N3, and N4 as three detection head layers D3, D2, and D1 of different sizes, respectively, to predict small, medium, and large target scales.

[0059] Please see Figure 4 The present invention also provides a remote sensing image aircraft detection system based on enhanced fine-grained features, characterized in that it includes: The preprocessing module is used to receive and preprocess the input remote sensing images; The detection module integrates an aircraft detection network constructed and trained according to a remote sensing image aircraft detection method based on enhanced fine-grained features as described above, for extracting features from the preprocessed image and detecting aircraft targets. The post-processing and output module is used to perform non-maximum suppression processing on the raw output of the detection module and generate the final detection result visualization information.

[0060] Training Example: The hardware environment was an NVIDIA GeForce RTX 2060 GPU. The software environment was PyTorch 1.3.0 and NumPy 2.2.6. Hyperparameter settings included a batch size of 8, 100 epochs, an initial learning rate of 0.01, and SGD optimizer (momentum = 0.9, weight decay = 5e-4). The training strategy used the CORS-ADD dataset, randomly partitioning the training, validation, and test sets in a 6:2:2 ratio. During training, the validation set was used to monitor model performance and save the optimal weights.

[0061] Detection Example: A 1024×1024 remote sensing image of an airport area is scaled down to 640×640 and then input into a pre-trained WS-YOLO model. The model outputs a series of predicted bounding boxes, which are then post-processed using non-maximum suppression (NMS). Finally, all detected aircraft targets are visualized on the original image, with their locations and confidence levels labeled.

[0062] The experimental results are shown in Table 1 below. The overall performance of the WS-YOLO model proposed in this invention is superior to five other comparative models. Specifically, the overall recall performance is improved by 1.74% to 7.14%, while the overall performance of mean precision (mAP) is improved by 0.44% to 4.95%. In terms of F1 score, the overall performance is improved by 0.59% to 5.00%. It can be seen that the proposed model improves the accuracy of aircraft detection without increasing the number of parameters, while maintaining a low computational cost of 6.8 GFLOPs and a high inference speed of 267 FPS.

[0063] Table 1. Comparison of the visualization detection results of the method of this invention with other comparative methods on the CORS-ADD dataset. Figure 5 Representative aircraft detection results are presented. Compared with five other models, the WS-YOLO model produces more accurate and comprehensive detection results. As shown in the first and fourth rows of the figure, the proposed model successfully detected small-scale aircraft targets missed by other models. In the second and third rows of the figure, the model demonstrates strong discriminative ability when faced with background clutter and visually similar ground objects. This advantage stems from the enhanced fine-grained module integrated into the WS-YOLO architecture. Overall, WS-YOLO demonstrates superior visual detection performance compared to all five comparison models, maintaining high accuracy and robustness even in scenes containing rich fine-grained aircraft details.

[0064] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for aircraft detection in remote sensing images based on enhanced fine-grained features, characterized in that, Includes the following steps: Step S1: Construct the WS-YOLO detection network; wherein the WS-YOLO detection network is a single-stage detector, comprising a backbone network, a neck network, and a detection head; the backbone network is divided into four extraction stages; Step S2: Construct an enhanced fine-grained module; wherein, the extraction stage uses convolutional layers, BN layers, and ReLU layers, and the enhanced fine-grained module is added after the ReLU layers of each extraction stage of the backbone network; Step S3: Use the remote sensing image dataset containing aircraft annotations to train the WS-YOLO detection network end-to-end to obtain the trained aircraft detection model; Step S4: Input the remote sensing image to be detected into the trained aircraft detection model, and output the bounding box position, category label and confidence score of the aircraft in the image; The processing flow of the enhanced fine-grained module includes: The input feature map is equally divided into first branch features along the channel dimension. Second branch features ; For the first branch feature Progressive fine-grained extraction is performed to obtain several intermediate features of the first branch; For the second branch features Perform channel importance recalibration to obtain the second branch calibration features; All intermediate features from the first branch and the calibrated features from the second branch are concatenated along the channel dimension to generate an enhanced fine-grained feature map.

2. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 1, characterized in that, The backbone network learns discriminative aircraft features and processes them sequentially through four extraction stages of downsampling and fine-grained enhancement modules, resulting in feature maps S1, S2, S3, and S4 in hierarchical order.

3. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 2, characterized in that, A context enhancement layer is also provided to process the deepest feature map S4 output by the backbone network. The feature map S4 is then processed by passing it through the SPPF module and C2PSA module in the YOLO model to generate high-level features Fs rich in global context information.

4. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 3, characterized in that, The neck network is used to perform multi-scale feature fusion and fine-grained enhancement, the specific process of which includes: The high-level features Fs, feature map S2, and feature map S3 are input into the neck network; The high-level feature Fs is continuously upsampled to obtain upsampled features at different scales; Multi-scale feature fusion is performed based on upsampled features at different scales, feature map S2 and feature map S3; Based on the fusion of aircraft features at different scales, cross-level information fusion features N1, N2, N3 and N4 are obtained.

5. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 4, characterized in that, The detection head is equipped with three detection head layers D1, D2 and D3 of different sizes corresponding to the three target scales of large, medium and small; wherein the fusion feature N2, fusion feature N3 and fusion feature N4 are respectively input to the detection head layer D3, detection head layer D2 and detection head layer D1.

6. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 5, characterized in that, The neck network inputs multi-scale fused features into the detection head, and the specific process includes: The high-level feature Fs is upsampled so that its spatial size is consistent with the feature map S3 and then fused to obtain the cross-level information fusion feature N1. The cross-level information fusion feature N1 is upsampled so that its spatial size is consistent with the feature map S2 and then fused to obtain the fused feature N2. The fusion feature N2 is upsampled so that its spatial size is consistent with the cross-level information fusion feature N1, and then fused to obtain the fusion feature N3. The high-level feature Fs and the fusion feature N3 are upsampled and fused to obtain the fusion feature N4; The fusion features N2, N3, and N4 are respectively input into the detection head layer D1, detection head layer D2, and detection head layer D3.

7. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 1, characterized in that, For the first branch feature The process of progressive fine-grained extraction includes: The first branch feature The input is fed into the RepConv layer to obtain the initial fine-grained features. ; The initial fine-grained features Input a densely connected block containing n-1 intermediate convolutional layers, and perform deep feature extraction to obtain n-1 intermediate features of the first branch; The last intermediate feature from the first branch is channel-fused using a 1×1 convolutional layer to obtain the final feature map. .

8. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 7, characterized in that, For the second branch features The channel importance recalibration process includes: For the second branch features Adaptive average pooling and adaptive max pooling are performed separately to aggregate spatial information; After adding the two pooling results, input them into a channel attention generation network consisting of a multilayer perceptron and a sigmoid activation function to generate a channel attention weight vector. The channel attention weight vector is multiplied channel by channel with the original second branch features to obtain the recalibrated features.

9. The method for aircraft detection in remote sensing images based on enhanced fine-grained features according to claim 1, characterized in that, The training process in step S3 uses a loss function that is a weighted sum of bounding box regression loss, classification loss, and distribution focal loss; wherein the bounding box regression loss is Wise-IoU Loss, the classification loss is BCE Loss, and the distribution focal loss is Distribution Focal Loss.

10. A remote sensing image aircraft detection system based on enhanced fine-grained features, characterized in that, include: The preprocessing module is used to receive and preprocess the input remote sensing images; The detection module integrates an aircraft detection network constructed and trained according to any one of claims 1 to 9, which is a remote sensing image aircraft detection method based on enhanced fine-grained features, for extracting features from the preprocessed image and detecting aircraft targets. The post-processing and output module is used to perform non-maximum suppression processing on the raw output of the detection module and generate the final detection result visualization information.

Images