SAR (Synthetic Aperture Radar) image aircraft target detection method based on multi-level collaborative deep network

Abstract

The invention discloses an SAR image aircraft target detection method based on a multi-level collaborative deep network, and the method comprises the steps: inputting an image into a trunk feature extraction network, firstly carrying out the preliminary feature extraction through a common convolution layer, and improving the image dimension while reducing the image size; then, through the synergistic effect of a feature extraction module based on explicit information adjustment and a global perception enhancement module based on Transform, deep global and local feature extraction is carried out, so that a feature map of feature aggregation is generated; next, designing a target strong scattering feature enhancement module based on adaptive collaborative attention in front of a detection head, and enabling the network to pay more attention to a target area through a direction weight distribution mechanism of the adaptive collaborative attention module, thereby further improving the feature expression capability; and finally, a collaborative optimization loss function fusing dynamic weight and precise positioning is designed, collaborative optimization of classification and positioning is realized, and thus the aircraft target detection performance under strong reflection background interference in the SAR image is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of SAR image aircraft target detection, and in particular to a SAR image aircraft target detection method based on a multi-level collaborative deep network. Background Art

[0002] Airport aircraft target detection technology provides core data support for military security, flight scheduling, and emergency rescue by sensing apron and runway dynamics in real time, making it a key component of modern aviation management systems. SAR, as an active microwave imaging technology, boasts all-day, all-weather imaging capabilities and can acquire high-resolution ground object information even under complex weather conditions, thus demonstrating significant advantages in target detection. However, due to the unique nature of SAR imaging mechanisms, aircraft target detection in SAR images faces numerous challenges. First, aircraft targets typically appear as a collection of discrete scattered points in SAR images, with uneven feature distribution and susceptible to target pose and imaging geometry. Second, coherent speckle noise in SAR images reduces the contrast between the target and the background, and the complex background of airport scenes can produce strong reflections similar to those of aircraft targets. Furthermore, in large-area scenes such as airports, the number of positive and negative samples is severely unbalanced, and aircraft targets are typically densely distributed, further increasing the detection difficulty.

[0003] In summary, aircraft target detection in SAR images is challenging due to the discrete target scattering points, strong scattering background interference, imbalance of positive and negative samples, and dense arrangement of targets. Therefore, it is necessary to develop new aircraft target detection methods in SAR images to overcome these difficulties and improve the accuracy and real-time performance of aircraft target detection in SAR images. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a SAR image aircraft target detection method based on a multi-level collaborative deep network.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] A method for detecting aircraft targets in SAR images based on a multi-level collaborative deep network includes the following steps:

[0007] Step 1: Based on the YOLOv5 network architecture, a global and local feature extraction network based on explicit information regulation is constructed. In the backbone feature extraction network, the EVC module and the Swin Transformer module are used to extract global and local features. After completing the backbone feature extraction step, the feature map will enter the YOLOv5 feature pyramid fusion network PAN.

[0008] Step 2: Build a target strong scattering feature enhancement network based on adaptive collaborative attention, add the adaptive collaborative attention module SCA between PAN and detection head, and further optimize the fused feature map;

[0009] Step three: Construct a collaborative optimization loss function that integrates dynamic weights and precise positioning, use weighted classification loss and regression loss as the final loss function, and finally, input the fused feature map into the detection head to obtain the detection result, completing the entire detection process.

[0010] It should be noted that the explicit information conditioning module EVC is introduced into the backbone feature extraction network. The module consists of a lightweight MLP and a learnable visual center LVC, including a smoothing block, deep convolution, channel MLP, and an intrinsic codebook.

[0011] It should be noted that the Swin Transformer module includes layer normalization, window area multi-head self-attention mechanism, shifted window area multi-head self-attention mechanism and multi-layer perceptron.

[0012] It should be noted that the backbone feature extraction network construction process is as follows: in the initial stage of feature extraction, continuous CBS (convolution-normalization-activation) and C3 modules are used to extract low-level features of the image. In the middle part of the network, the EVC module is introduced after the second C3 module, and the feature extraction capability is collaboratively optimized through the lightweight MLP and LVC mechanisms. In the second half of the network, the Swin Transformer module is used to replace the neck structure of the C3 module to form a new STRC3 module. Finally, the feature map is input into the SPPF network to complete the construction of the backbone feature extraction network.

[0013] It should be noted that the adaptive collaborative attention module is based on the CA attention structure and obtains a feature map with attention weights by introducing a learnable weight generation network.

[0014] It should be noted that the adaptive collaborative attention module is embedded between the PAN and the detection head, which can further optimize the fused feature map.

[0015] It should be noted that the collaborative optimization loss function consists of two parts: Varifocal Loss (VFL) classification loss and CIoULoss regression loss. The two loss functions are jointly weighted as the overall loss of the detection algorithm.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0017] 1. In order to solve the problem that the scattering points of aircraft targets in SAR images are discretely distributed, which makes it difficult to extract complete aircraft features, the present invention constructs a global and local feature extraction network based on explicit information regulation. In the backbone feature extraction network, an explicit visual center module is used, and its deep convolution is used to perform sliding window processing in the spatial dimension to effectively extract the global features of the SAR image. At the same time, the input features are matched with the templates in the inherent codebook to extract local information related to the aircraft target in the image, thereby enhancing the model's attention to local key areas. At the same time, the self-attention mechanism of the Transformer is used to calculate the correlation between image pixels and aggregate the information of discrete strong scattering points, thereby improving the overall expression of aircraft targets in SAR images.

[0018] 2. In SAR imagery of airport scenes, since a large number of artificial structures such as runways and aprons produce strong reflection features similar to those of aircraft targets, leading to false alarms during detection, an algorithm for enhancing target strong scattering features based on adaptive collaborative attention is constructed. This algorithm is based on coordinate attention and innovatively introduces a directional weight adaptive allocation module. By performing average pooling analysis on the target's horizontal and vertical scattering intensities, it calculates directional weights and dynamically adjusts the focus of the target area in the feature map. This design not only accurately captures the target's strong scattering features in different directions, but also effectively suppresses background interference, thereby significantly improving the detection performance of aircraft targets in complex background environments.

[0019] 3. To address the issues of aircraft targets occupying a small pixel area in airport SAR images, resulting in an imbalance between positive and negative samples and prone to false alarms, as well as the reduced positioning accuracy caused by densely packed targets, a collaborative optimization loss function for steel components is developed that integrates dynamic weighting and precise positioning. This loss function combines the dynamic weight classification advantages of Varifocal Loss with the precise positioning regression advantages of CIoULoss, allowing the model to focus more on high-quality positive samples, improving classification accuracy. Furthermore, it can more accurately fit target boundaries in dense target scenes, reducing missed detections and false positives, significantly improving the model's detection performance in these dense target scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of the overall network architecture of the present invention;

[0021] Figure 2 Schematic diagram of the structure of the explicit information adjustment module of the present invention;

[0022] Figure 3 Schematic diagram of the Swin Transformer module structure of the present invention;

[0023] Figure 4 Schematic diagram of the STRC3 module structure of the present invention;

[0024] Figure 5 This is a schematic diagram of the backbone feature extraction network structure of the present invention;

[0025] Figure 6 Schematic diagram of the CA attention structure of the present invention;

[0026] Figure 7 Schematic diagram of the structure of the adaptive collaborative attention module of the present invention;

[0027] Figure 8 Schematic diagram of the target strong scattering feature enhancement network structure based on adaptive collaborative attention in the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the following embodiments are based on the present technical solution and provide detailed implementation methods and specific operating processes, but the protection scope of the present invention is not limited to these embodiments.

[0029] like Figures 1 to 8 As shown, the present invention is a method for detecting aircraft targets in SAR images based on a multi-level collaborative deep network, comprising the following steps:

[0030] Step 1: Based on the YOLOv5 network architecture, a global and local feature extraction network based on explicit information regulation is constructed. In the backbone feature extraction network, the EVC module and the Swin Transformer module are used to extract global and local features. After completing the backbone feature extraction step, the feature map will enter the YOLOv5 feature pyramid fusion network PAN.

[0031] Step 2: Build a target strong scattering feature enhancement network based on adaptive collaborative attention, add the adaptive collaborative attention module SCA between PAN and detection head, and further optimize the fused feature map;

[0032] Step three: Construct a collaborative optimization loss function that integrates dynamic weights and precise positioning, use weighted classification loss and regression loss as the final loss function, and finally, input the fused feature map into the detection head to obtain the detection result, completing the entire detection process.

[0033] like Figure 1 As shown in the figure, the present invention uses YOLOv5 as the basic network architecture. First, the EVC module is introduced into the backbone feature extraction network to extract global and local features. The specific module framework is shown in Figure 2As shown. This module consists of a parallel lightweight MLP and a learnable visual center LVC. The feature map obtained by C3 does not directly enter the parallel MLP and LVC. It will first be processed by a smoothing block for smoothing features. It consists of a 7×7 convolution, a BN layer, and a ReLu function. The smoothing block can reduce noise while accelerating calculations and improving visual interpretability. After processing by the smoothing block, the feature map enters the parallel block for deeper feature extraction. Finally, the features extracted by the two parallel branches are spliced along the channel dimension to obtain the final output result. This process can be expressed as follows:

[0034] Z=Concat(MLP(X stem ), LVC(X stem ))

[0035] The EVC module integrates upper and lower branch features through smoothing blocks and utilizes a lightweight MLP (including deep convolution and channel-wise MLP residual modules) to extract global contextual information and enhance the correlation of discrete scattering points. Deep convolution expands the receptive field, channel-wise MLP optimizes feature weights, and residual connections preserve key information. The LVC module dynamically weights local features (such as wing and fuselage scattering areas) using an inherent codebook and a learnable scaling factor to highlight key target areas. The combination of these two, complementing global and local features, improves the detection accuracy of aircraft targets in SAR imagery.

[0036] Specifically, the Swin Transformer module is introduced into the backbone feature extraction network to enhance the modeling ability of global information. This module is used to replace the neck structure of the C3 module to form a new STRC3 module. The specific module framework is shown in the figure below. Figure 4 The Swin Transformer module structure is shown in Figure 3 As shown in the figure, it mainly consists of layer normalization, multi-layer perceptron, window area multi-head self-attention and shifted window area multi-head self-attention. Window area multi-head self-attention and shifted window area multi-head self-attention are two key multi-head self-attention modules in Swin Transformer, which are used to capture local information and model long-distance dependencies respectively. In traditional Transformer, the self-attention mechanism needs to be calculated for all positions of the entire image, which results in computational complexity proportional to the square of the image size. The amount of computation is huge when processing large-size images. Swin Transformer significantly reduces computational complexity by dividing the image into multiple non-overlapping windows and limiting the self-attention calculation within each window. The calculation process is as follows:

[0037]

[0038] Specifically, the backbone feature extraction network construction process is as follows: In the initial feature extraction stage, a continuous CBS (Convolution-Normalization-Activation) and C3 module are used to capture low-level features of SAR images. This enhances the ability to capture edge and texture information, improves the feature resolution of key aircraft components, and lays the foundation for subsequent feature representation. In the middle of the network, the EVC module is introduced to collaboratively optimize feature extraction capabilities through a lightweight MLP and LVC mechanism. The lightweight MLP is connected to a channel-wise MLP via a deep convolutional module to construct a global feature representation, enhancing the network's understanding of the target's overall morphology and alleviating the problem of discrete target scattering points in SAR images. The learnable visual center module LVC uses a codebook matching mechanism to match input features with predefined codebook templates to extract local salient features related to the target, targeting the non-uniform scattering characteristics of aircraft targets in SAR images. This mechanism adaptively enhances key features such as strong scattering points and edge structures in the target area while effectively suppressing background interference, thereby improving the network's ability to extract local features of aircraft targets. In the second half of the network, the STRC3 module is introduced, and the sliding window self-attention mechanism of the Swin Transformer is used to model long-distance dependencies to further integrate the feature information of each part of the target and enhance the expression ability of the global features. Finally, the feature map is input into the SPPF network, which extracts the multi-scale features of the target through three pooling operations and performs splicing and fusion to expand the receptive field, enabling the network to more comprehensively understand the structural information of the target at different scales. The specific backbone feature extraction network structure diagram is shown in the figure below. Figure 5 shown.

[0039] Specifically, the adaptive collaborative attention module is based on the CA attention structure and obtains a feature map with attention weights by introducing a learnable weight generation network. The network structure diagram of the CA attention mechanism is shown in the figure below. Figure 6As shown in Figure 2, the CA attention mechanism is the foundation of adaptive collaborative attention. This attention mechanism considers both channel information and coordinate position information, enhancing the saliency of target regions in feature maps. The CA attention mechanism performs global average pooling on the input feature map in both the horizontal and vertical directions to extract feature information in each direction. The two directional pooling results are then concatenated to produce a feature map with rich spatial features. The concatenated feature map is then processed through a batch normalization (BN) layer. The BN layer normalizes the data distribution, reduces internal drift during training, accelerates network convergence, and stabilizes the training process. Next, the normalized feature map is activated using a sigmoid function in both the horizontal and vertical directions, converting the feature maps in both directions into probability values. The sigmoid operation compresses each pixel value in the feature map to a value between 0 and 1, indicating the degree of attention paid to the feature in that direction at each location. Finally, this feature map is weighted with the original input feature map to produce a feature map with attention weights. In SAR images, background regions often contain strong reflection signals that are similar to target features and can easily interfere with aircraft detection. Through weighted calculation, the network can increase the weight of the aircraft target area according to the feature strength of the target area and reduce the influence of background information, thereby improving the detection accuracy of aircraft targets and reducing the false alarm rate of background detection.

[0040] CA attention can combine spatial information on the basis of channel attention, so that the network can focus on more directional features. However, it processes information in different directions in a relatively balanced way during the calculation process, and cannot effectively distinguish the difference in scattering intensity of aircraft targets in different directions in SAR images. To solve this problem, the present invention designs a self-adaptive coordinated attention module (SCA), such as Figure 7 As shown in Figure 1, the SCA module uses a learnable weight generation network based on CA attention. This paper uses a fully connected layer to dynamically calculate the weight distribution coefficients in two directions. The weight distribution function can be expressed as:

[0041] α x =σ(W x ·f x +b x ), α y =σ(W y ·f y +b y )

[0042] W x =W x1 W x2 W x3 , Wy =W y1 W y2 W y3

[0043] Among them, f x represents the normalized feature map, W x is the weight matrix, σ(·) is the activation function, b x is the bias vector. The fully connected layer linearly transforms the feature map through a set of weight matrices and bias vectors and generates a new feature representation α x . In this process, each input node (pixel feature value) is multiplied by the weight matrix of the fully connected layer, weighted summed by adding the bias term, and then a new feature representation is obtained through the activation function. This new feature representation is used to adjust the feature weights in a specific direction, so that the network pays more attention to target areas with stronger scattering in different directions. For example, in the wing part of an aircraft, due to its high scattering intensity in the horizontal direction, the adaptive weight mechanism will allocate more attention to the horizontal direction, enhancing the feature expression of the wing part in the horizontal direction. In this way, the adaptive collaborative attention module can enhance the strong scattering features of aircraft targets in different directions, effectively improving the accuracy of aircraft target detection.

[0044] Specifically, the adaptive collaborative attention module is added after the PAN feature fusion network of YOLOv5 and before the detection head, as shown in Figure 8 As shown in the figure, a target feature enhancement network based on adaptive allocation of directional weights is constructed to further optimize the fused feature maps. The feature fusion network is an important component of the YOLO series of algorithms. Its function is to deeply fuse multiple feature maps of different scales output by the backbone feature extraction network to improve the prediction accuracy of the detection network. The feature fusion network of YOLOv5 adopts the PAN feature pyramid structure. During the feature fusion process, the algorithm first performs a bottom-up operation, sequentially upsampling the feature maps so that the low-level feature maps can gradually match the high-level feature maps in size, thereby achieving effective information transmission and fusion; then, the feature maps are downsampled from top to bottom to ensure that the size of each layer of feature maps ultimately remains consistent. In this way, the feature information of each layer can be effectively fused, and the advantages of feature maps of different scales can be fully utilized.

[0045] During the feature fusion stage, low-level feature maps are upsampled and then fused with high-level feature maps. Although upsampling can restore the spatial resolution of the feature map, the detail information contained in the feature map itself is relatively coarse. The upsampled information may be blurred and unable to fully restore the fine structure of the target. This process will weaken the detailed features of the target area. Embedding the SCA module between the PAN and the detection head can further optimize the fused feature map. SCA adaptively adjusts the weights of each area in the feature map based on the scattering characteristics of the target in different directions, ensuring that key target areas such as the aircraft's wings and engines have stronger feature expression in the final output, reducing background noise interference, thereby making the target features more prominent and improving the model's sensitivity to the target area.

[0046] Specifically, the collaborative optimization loss function is jointly weighted by Varifocal Loss and CIoU Loss as the overall loss of the proposed algorithm. VFL introduces a dynamic weight mechanism that can adjust the weight according to the difference between the predicted probability of the sample and the true label. For foreground samples that are easily misclassified as background, that is, high-quality positive samples, Varifocal Loss will give them higher weights, so that the model pays more attention to these samples during the training process, thereby improving the detection ability of aircraft targets. In SAR images, since aircraft targets may be affected by coherent speckle noise, complex background and other factors, the features of some aircraft targets may be similar to the background and are easily misclassified as background. VarifocalLoss uses a dynamic weight mechanism to enable the model to pay more attention to these difficult-to-classify aircraft target samples, strengthen the learning of their features, and thus improve the accuracy of classification. The expression of Varifocal Loss is:

[0047]

[0048] CIoU Loss introduces three comprehensive constraints: center point distance, aspect ratio, and intersection over union (IoU). Compared to the traditional IoU loss function, CIoU Loss not only focuses on the overlap between the predicted box and the ground-truth box, but also considers the center distance and aspect ratio of the target box. Therefore, it can more accurately measure the similarity between two bounding boxes. During bounding box regression, if the center point distance between the predicted box and the ground-truth box is large, even if their IoU ratio is high, the aircraft target may still not be accurately located. By calculating the center point distance, CIoU Loss encourages the predicted box to move closer to the ground-truth box more quickly, guiding the model to learn more accurate target location information. In SAR images, the poses of aircraft targets vary greatly, resulting in varying aspect ratios within the image. CIoU Loss effectively optimizes the shape and position of the target box by adjusting the aspect ratio of the predicted box, reducing box overlap between densely packed objects, improving the accuracy of the target boundary, and thus reducing the incidence of multiple detections and missed detections. The calculation formula of the CIoU Loss function is as follows:

[0049]

[0050] Next, the two loss functions are jointly weighted as the overall loss of the proposed algorithm. The collaborative optimization loss function is:

[0051] L=w1L VFL +w2L CIoU

[0052]

[0053] w1 and w2 are the dynamic weights for the VFL and CIoU Loss, respectively. λ is a balancing factor that controls the dynamic magnitude of the loss terms. Compared to fixed weight coefficients, this design adaptively adjusts weights based on the magnitude of the loss. When a particular loss is large, its weight is automatically reduced, prompting the network to focus more on relatively small losses during training, making the training process more stable and preventing a single loss from dominating the optimization process. The weights are calculated using a softmax method, ensuring that the weights of the two loss terms always remain between [0, 1] and sum to 1, ensuring the stability of loss optimization. During training, if one loss term is significantly larger than the other, the smaller loss term is given a greater weight, prompting the model to learn the classification and regression tasks more balanced, improving overall detection performance.

[0054] When the present invention is working, the image is input into the backbone feature extraction network, and low-level feature extraction is first performed through the continuous CBS (convolution-normalization-activation) and C3 modules in YOLOv5. Then, the EVC module and the SwinTransformer module are used in the backbone network to perform global and local fine feature extraction to generate a feature map of feature aggregation. After completing the backbone feature extraction step, the feature map will enter the feature pyramid fusion network PAN of YOLOv5 to generate feature maps of different scales, and then enter the target strong scattering feature enhancement network based on adaptive collaborative attention to further optimize the fused feature map. A collaborative optimization loss function that fuses dynamic weights and precise positioning is constructed at the regression of the detection head, and the weighted classification loss and regression loss are used as the final loss function. Finally, the fused feature map is input into the detection head to obtain the detection result, completing the entire detection process.

[0055] Those skilled in the art can make various corresponding changes and modifications based on the above technical solutions and concepts, and all of these changes and modifications should be included in the scope of protection of the claims of the present invention.

Claims

1. A method for detecting aircraft targets in SAR images based on a multi-level collaborative deep network, characterized in that: The following steps are involved: Step 1: Using YOLOv5 as the basic network architecture, we build a global and local feature extraction network based on explicit information conditioning. In the backbone feature extraction network, we introduce the explicit information conditioning module (EVC) to achieve the coordinated extraction and conditioning of global and local features. At the same time, we introduce the Swin Transformer module to further enhance the global perception capability of the network and utilize its sliding window self-attention mechanism to efficiently capture the global dependencies of the image. Step 2: Build an algorithm to enhance the strong scattering features of targets based on adaptive collaborative attention. Through the directional adaptive weight mechanism in the adaptive collaborative attention module, the network's attention to target features is improved, while background information is weakened, significantly enhancing the target's feature expression. Step 3: In the post-processing stage, a collaborative optimization loss function that integrates dynamic weights and precise positioning is introduced into the loss function of bounding box regression to achieve collaborative optimization of classification and positioning regression.

2. The SAR image aircraft target detection method based on multi-level collaborative deep network according to claim 1 is characterized in that: The explicit information conditioning module EVC is first introduced into the backbone feature extraction network in step 1. The module consists of a lightweight MLP and a learnable visual center LVC, including a smoothing block, a depth convolution, a channel MLP, and an intrinsic codebook.

3. The SAR image aircraft target detection method based on a multi-level collaborative deep network according to claim 1 or 2 is characterized in that: In the step 1, the Swin Transformer module is introduced into the backbone feature extraction network, and the module includes layer normalization, window area multi-head self-attention mechanism, shifted window area multi-head self-attention mechanism and multi-layer perceptron.

4. The method for detecting aircraft targets in SAR images based on a multi-level collaborative deep network according to claim 3, wherein: In step 1, the backbone feature extraction network construction process is as follows: In the initial feature extraction phase, sequential CBS and C3 modules are used to extract low-level features from SAR images. In the middle of the network, an EVC module is introduced after the second C3 module, using a lightweight MLP and LVC mechanism to collaboratively optimize feature extraction capabilities. In the latter half of the network, a Swin Transformer module replaces the neck structure of the C3 module, forming a new STRC3 module. Finally, the feature map is input into the SPPF network.

5. The target strong scattering feature enhancement based on adaptive collaborative attention according to claim 1, characterized in that: In the step 2, the adaptive collaborative attention module is based on the CA attention structure and obtains a feature map with attention weights by introducing a learnable weight generation network.

6. The target strong scattering feature enhancement based on adaptive collaborative attention according to claim 1 or 5, characterized in that: In the step 2, the adaptive collaborative attention module is added after the PAN feature fusion network of YOLOv5 and before the detection head, which can further optimize the fused feature map.

7. The collaborative optimization loss function integrating dynamic weight and precise positioning according to claim 6 is characterized in that: In step 3, the collaborative optimization loss function is jointly weighted by Varifocal Loss and CIoU Loss as the overall loss of the proposed algorithm.

Images