Method for monitoring ship and aircraft based on ArgusSAR (Synthetic Aperture Radar)

Abstract

The invention discloses a method for monitoring ships and aircrafts based on ArgusSAR, and belongs to the field of remote sensing image processing, and the method comprises the steps: obtaining ship and aircraft related SAR images, and obtaining a training set and a test set through marking and dividing; an ArgusSAR model containing a space encoder, a fine-grained feature aggregation encoder and a multi-dimensional perception attention decoder is constructed, and local and global feature extraction, multi-scale feature fusion and scale-space-task perception are realized respectively; and based on the training set and the multi-task joint loss function training model, inputting the test set into the optimal model to output the target position and category. The method is high in background interference resistance, can accurately detect dense and shielded targets, is adaptive to multi-scale targets, and is stable in cross-dataset generalization. Three data set tests show that core indexes are superior to mainstream algorithms such as a YOLO series, and a high-precision solution is provided for scenes such as maritime affair supervision and aviation safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and specifically to a method for monitoring ships and aircraft based on ArgusSAR. Background Technology

[0002] Because of the significant difference in backscattering coefficients between metallic materials and the ground / sea surface background in SAR (Synthetic Aperture Radar) images, targets such as ships and aircraft can be effectively distinguished from other interfering objects (such as buildings, sea clutter, etc.), thus finding widespread application in military and civilian target detection. Currently, target detectors face challenges such as complex background interference, dense target arrangement, and weak multi-scale target generalization ability. Their performance significantly decreases when detecting different SAR image bands and different ground objects, and their generalization ability struggles to meet practical application requirements. Summary of the Invention

[0003] The purpose of this invention is to propose a method for monitoring ships and aircraft based on ArgusSAR, so as to solve the problems of reduced detection performance and weak generalization ability in the prior art.

[0004] The technical solution adopted in this invention is: a method for monitoring ships and aircraft based on ArgusSAR, including the following steps: Step S1, acquiring SAR image data related to ships and aircraft, performing multi-target annotation and dataset division on the SAR image data, and obtaining a multi-target SAR image dataset containing training set and test set.

[0005] Step S2: Construct the ArgusSAR model using a spatial encoder, a fine-grained feature aggregation encoder, and a multi-dimensional perceptual attention decoder. The spatial encoder includes a convolutional coding module and a multi-head attention module, used to extract local and global interaction features from the feature map. The fine-grained feature aggregation encoder includes a convolutional kernel prediction module and a content-aware reconstruction module, used to achieve hierarchical multi-scale feature fusion. The multi-dimensional perceptual attention decoder includes scale-aware attention, spatial-aware attention, and task-aware attention, used to achieve scale perception, spatial perception, and task perception of the target.

[0006] Step S3: Train the constructed ArgusSAR model based on the training set, optimize the model parameters through a multi-task joint loss function, and obtain the ArgusSAR model with the best recognition performance.

[0007] Step S4: Input the test set into the ArgusSAR model with the best recognition effect, and use the model to detect ship and aircraft targets in SAR images, and output the location and category recognition results of the target objects.

[0008] As a further improvement of the present invention, in step S2, the specific working method of the spatial encoder includes: step S211, input embedding: the input features are represented as an embedding matrix. ,in It is the sequence length. It is the embedded dimension.

[0009] Step S212, Linear Transformation: For each attention head Embedding matrix After input, the matrix is ​​processed with three different weight matrices, and a three-step linear transformation is performed to obtain the query matrix. Key matrix Sum matrix The calculation formula is as follows:

[0010] ,in, , and Is with the first The weight matrix associated with each attention head.

[0011] Step S213, Attention Score Calculation: When calculating the attention score for different channels and spatial locations, for each attention head... First, calculate the query matrix. Bond matrix The dot product between them, divided by the key dimension The square root of the matrix is ​​scaled, and then the attention score matrix is ​​obtained by applying the Softmax function. The calculation formula is as follows: .

[0012] Step S214, Weighted Summation: Using the attention score matrix Log-value matrix Perform a weighted sum operation to obtain the first... The output of each attention head The calculation formula is as follows: .

[0013] Step S215, splicing transformation: obtaining the attention head After outputting the values, all output matrices are concatenated, and a linear transformation is performed on the concatenated output matrix to obtain the final weighted embedding representation. The calculation formula is as follows: ,in, Wo This is the initial random weight matrix.

[0014] As a further improvement of the present invention, in step S2, the specific working method of the fine-grained feature aggregation encoder includes: step S221, based on the content of the input low-resolution feature map, the convolution kernel prediction module dynamically generates a specific upsampling kernel for subsequent processes according to the SAR image content, specifically: step S2211, for the input features... After channel compression, the result is obtained Through a convolution kernel of size Convolutional layer processing features The number of output channels is σ is the upsampling factor. It is the content-aware kernel size.

[0015] Step S2212, these Unfold in space to provide a high-resolution image of the target at every location. The generated size is The upsampling kernel.

[0016] Step S2213: Perform Softmax normalization independently on each generated upsampling kernel to ensure that its weight sum is 1, and obtain the final upsampling kernel used to guide the content-aware reconstruction module.

[0017] Step S222: For each target location on the output feature map The convolutional kernel prediction module first locates and extracts a kernel of size [size missing] on its corresponding low-resolution input feature map. The local neighborhood is used to obtain the predicted size of the target location. A specific upsampling kernel is used; the features extracted from the low-resolution neighborhood are weighted and summed using the upsampling kernel to obtain the final high-resolution fused feature map. .

[0018] As a further improvement of the present invention, in step S2, the specific working method of the multidimensional perceptual attention decoder includes: step S231, feature alignment and stacking: receiving multi-level feature maps from the fine-grained feature aggregation encoder. All feature maps are aligned by adjusting their size and number of channels, and then stacked into a uniform feature tensor. This is to prepare for subsequent processing.

[0019] Step S232, Scale-aware attention processing: The feature tensor... Input to scale-aware attention module πl The scale-aware attention module calculates the weights of different feature levels using the following formula and weights the features accordingly to dynamically fuse features at different scales, enhancing adaptability to targets of different sizes and obtaining scale-enhanced features. Fπl : Where C represents the number of channels and S represents the number of feature scales. It is a linear function. It is an activation function.

[0020] Step S233, Spatial Perception Attention Processing: This involves processing the features that have undergone scale perception... F πl Input to spatial awareness attention module πs The spatial awareness attention module learns spatial offsets through self-learning. Δ pk and importance scalar Δ mk Feature aggregation is performed on sparse sampling points, enabling the model to focus on key spatial regions of the target, suppressing interference, and obtaining spatially enhanced features. F πs ： ,in, L Represents the number of feature levels. K It is the number of sparse sampling locations. pk It is a spatial scalar.

[0021] Step S234, Task-Aware Attention Processing: Processing the features that have undergone spatial perception... F πs Input to the task-aware attention module πc The task-aware attention module adaptively allocates and optimizes feature channels for different sub-tasks of classification and localization through a gating function, enhancing the decoder's adaptability to multiple tasks and ultimately generating feature maps for object detection. :

[0022] Where α and β are adaptive parameters, and It is a gating function that learns to control the activation threshold. First of all Global average pooling is performed along the dimension, followed by two fully connected layers and a normalization layer. Finally, the sigmoid function is applied to normalize the output. scope; F πs,c It is the feature slice of the c-th channel.

[0023] As a further improvement of the present invention, in step S3, the multi-task joint loss function is composed of localization loss. Confidence loss and category loss Composition, defined as follows:

[0024] Among them, (1) positioning loss The calculation formula is:

[0025] ;in, This represents the square of the distance between the center points of the ground truth bounding box and the predicted bounding box, while This represents the diagonal length of the smallest bounding rectangle that can contain both boxes. and The calculation formula is as follows: , IOU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. ω gt It is the width of the actual frame. It is the width of the prediction box. h gt It is the actual height of the frame. h It is the height of the predicted bounding box.

[0026] (2) Confidence loss The calculation formula is:

[0027] ,in, This indicates that a target exists. This indicates that there is no target. These represent the balancing weights for the loss caused by negative samples, used to reduce the contribution of a large number of negative samples to the loss function. This represents the bounding box predicted by the model. s' represents the actual bounding box, while s' represents the number of rows and columns of the image.

[0028] (3) Category loss The calculation formula is:

[0029] ,in, The category label representing the target object. This represents the probability predicted by the model.

[0030] Compared with the prior art, the present invention has the following technical advantages:

[0031] The ArgusSAR-based method for monitoring ships and aircraft of this invention exhibits strong robustness in SAR ship and aircraft object detection, making it more suitable for practical applications. It provides a valuable solution for the application of high-precision target detectors, specifically:

[0032] (1) High generalization ability: Based on joint training of multi-source SAR datasets (SSDD, HRSID, SADD) and optimization of multi-task loss function, the performance is stable across sensor and target type (ship / aircraft) scenarios.

[0033] (2) Leading performance: On the three major datasets, the core indicators such as P, R, F1, and mAP are all superior to SSD, YOLO series and various improved algorithms. The comprehensive detection accuracy and robustness reach the advanced level in the industry, providing highly practical solutions for scenarios such as maritime supervision and aviation case monitoring.

[0034] (3) Strong anti-interference capability: The spatial encoder’s local and global feature interaction and multi-dimensional perception attention spatial sparse sampling can accurately suppress background interference such as sea surface clutter and buildings, and significantly reduce false positives and false negatives.

[0035] (4) Accurate detection of dense targets: The content-aware reconstruction of the fine-grained feature aggregation encoder integrates high and low resolution features to solve the feature confusion problem caused by target occlusion and dense arrangement, and the false detection rate is close to zero.

[0036] (5) Full-scale adaptation: Scale-aware attention dynamically adjusts the feature weights of different resolutions to adapt to size differences from small aircraft to large ships, covering the needs of multi-scale target detection. Attached Figure Description

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Figure 1 This is a schematic diagram of the ArgusSAR model framework of the present invention.

[0039] Figure 2 This is a diagram of a multi-head self-attention structure.

[0040] Figure 3 This is a framework diagram of a fine-grained feature aggregation encoder.

[0041] Figure 4 This is a framework diagram of a multidimensional perceptual attention decoder.

[0042] Figure 5 The visualization compares the results of the SSDD dataset. The green bounding box represents TP, the blue bounding box represents FP, and the red bounding box represents FN.

[0043] Figure 6 This is a visualization comparing the results of the HRSID dataset. The green bounding box represents TP, the blue bounding box represents FP, and the red bounding box represents FN.

[0044] Figure 7This is a visualization comparing the results of the SADD dataset. The green bounding box represents TP, and the blue bounding box represents FP.

[0045] Figure 8 The first row shows the ArgusSAR results, and the second row shows the YOLOv10 results, which are used to visualize the attention heatmap of the detector. Detailed Implementation

[0046] 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. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0047] like Figure 1 As shown, the present invention discloses a method for monitoring ships and aircraft based on ArgusSAR, including the following steps: Step S1, acquiring SAR image data related to ships and aircraft, performing multi-target annotation and dataset division on the SAR image data, and obtaining a multi-target SAR image dataset containing training set and test set.

[0048] This embodiment uses three different datasets: the SAR Ship Detection Dataset (SSDD), the High-Resolution SAR Images Dataset (HRSID), and the SAR Aircraft Detection Dataset (SADD). The SSDD dataset contains 1160 SAR images from different sensors, including RadarSat-2, TerraSAR-X, and Sentinel-1. The HRSID dataset contains 136 panoramic SAR images from different sensors, including TerraSAR-X, TanDEM-X, and Sentinel-1, with a total of 16951 high-quality labeled ship targets. The SADD dataset contains 2966 non-overlapping images (224 × 224 pixels) collected by the TerraSAR-X satellite, containing 7835 aircraft targets. To ensure sufficient model training and objective testing, the three datasets are randomly partitioned in a 7:3 ratio, meaning 70% of the image samples (including corresponding annotations) in each dataset are used as the training set for model parameter learning, and the remaining 30% are used for the training set. The image samples are used as the test set to evaluate the detection performance of the model. During the partitioning process, it is ensured that the sample distribution of the training set and the test set is consistent to avoid the reliability of the model training and testing results being affected by the data distribution deviation.

[0049] Step S2: Construct the ArgusSAR model using a spatial encoder, a fine-grained feature aggregation encoder, and a multi-dimensional perceptual attention decoder. The spatial encoder includes a convolutional coding module and a multi-head attention module to extract local and global interaction features from the feature map. The fine-grained feature aggregation encoder includes a convolutional kernel prediction module and a content-aware reconstruction module to achieve hierarchical multi-scale feature fusion. The multi-dimensional perceptual attention decoder includes scale-aware attention, spatial-aware attention, and task-aware attention to achieve scale perception, spatial perception, and task perception of the target.

[0050] Specifically, the spatial encoder comprises two modules: a multi-head attention module and a convolutional coding module. By enhancing the interaction between local and global features, it focuses more on the important features of ships and aircraft objects and suppresses interference from irrelevant features. The workflow of the spatial encoder is as follows:

[0051] Step 1, Input Embedding: The input features are represented as an embedding matrix. ,in It is the sequence length. It is the embedded dimension.

[0052] Step 2, Linear Transformation: For each attention head Embedding matrix After input, the matrix is ​​processed with three different weight matrices, and a three-step linear transformation is performed to obtain the query matrix. Key matrix Sum matrix The calculation formula is as follows:

[0053] ,in, , and Is with the first The weight matrix associated with each attention head.

[0054] Step 3, Attention Score Calculation: When calculating the attention score for different channels and spatial locations, for each attention head... First, calculate the query matrix. Bond matrix The dot product between them, divided by the key dimension The square root of the matrix is ​​scaled, and then the attention score matrix is ​​obtained by applying the Softmax function. The calculation formula is as follows:

[0055] .

[0056] Step 4, Weighted Summation: Using the attention score matrix Log-value matrix Perform a weighted sum operation to obtain the first... The output of each attention head The calculation formula is as follows:

[0057] .

[0058] Step 5: Splicing and Transformation: Obtaining the Attention Head After outputting the values, all output matrices are concatenated, and a linear transformation is performed on the concatenated output matrix to obtain the final weighted embedding representation. The calculation formula is as follows:

[0059] .

[0060] Through the above operations, multi-head self-attention structures (such as...) Figure 2 As shown, different regions are weighted, enabling the model to dynamically adjust the relative importance of different channels or regions. This effectively suppresses irrelevant features of interfering objects (such as buildings, ground facilities, etc.) and ensures that the model can focus on extracting accurate and stable feature representations of ship and aircraft objects.

[0061] Specifically, the fine-grained feature aggregation encoder adopts a multi-level pyramid structure and uses a hierarchical design to integrate multi-scale feature information. Figure 3 Here is its structural diagram. The workflow of the fine-grained feature aggregation encoder is as follows:

[0062] Step 1: Based on the input low-resolution feature map content, the convolution kernel prediction module dynamically generates specific upsampling kernels for subsequent processes according to the SAR image content.

[0063] First, the input features After channel compression, the result is obtained After passing through a convolution kernel of size Convolutional layer processing features , The size of the kernel determines the range of the local context receptive field that the module can utilize when predicting the upsampling kernel, and the number of output channels is... .

[0064] Next, these predicted weights are spatially expanded to represent each location in the target high-resolution image. The generated size is The upsampling kernel.

[0065] Finally, each generated upsampling kernel is independently Softmax normalized to ensure that its weight sum is 1, resulting in the final upsampling kernel used to guide the content-aware reconstruction module.

[0066] Step 2: Receive the original low-resolution feature map and the upsampling kernel generated by the convolution kernel prediction module for each high-resolution output location as input.

[0067] For each target location on the output feature map This module first locates and extracts a feature map of size [size missing] on its corresponding low-resolution input feature map. The local neighborhood.

[0068] Next, it obtains the predicted size for the target location. Specific upsampling kernels.

[0069] Finally, the features extracted from the low-resolution neighborhood (all channels share the same kernel) are weighted and summed using this upsampling kernel, and the result is the final high-resolution fused feature map. .

[0070] The size needs to be consistent with To achieve the best results, the invention was used in subsequent experiments to verify the optimal parameters for ship and aircraft target detection in SAR images.

[0071] Specifically, the multi-dimensional perceptual attention decoder effectively combines multiple self-attention mechanisms to achieve the fusion of multi-scale, multi-target, and multi-task detection decoders. Figure 4 Here is its structural diagram. The structure of the multidimensional perceptual attention decoder includes scale-aware attention. πl Spatial perception and attention πs and task-aware attention πc Its workflow is as follows:

[0072] Step 1, Feature Alignment and Stacking: Receive multi-level feature maps from the fine-grained feature aggregation encoder. All feature maps are aligned by adjusting their size and number of channels, and then stacked into a uniform feature tensor. This is to prepare for subsequent processing.

[0073] Step 2, Scale-aware attention processing: The feature tensor... Input scale-aware attention module π l The scale-aware attention module calculates the weights of different feature levels using the following formula and weights the features accordingly to dynamically fuse features at different scales, enhancing adaptability to targets of different sizes and obtaining scale-enhanced features. F πl : Where C represents the number of channels and S represents the number of feature scales. It is a linear function. It is an activation function. πl It is a scale-aware attention mechanism that dynamically fuses features from different scales to enhance scale invariance. It learns the relative importance of different feature levels by adaptively adjusting the information contribution from feature maps of different resolutions, based on the semantic importance of the target (implicitly related to its scale). For example, it may emphasize high-resolution features for small objects and low-resolution but semantically stronger features for large objects, thus smoothing out scale differences between different levels.

[0074] Step 3, Spatial Perception Attention Processing: This involves processing features that have undergone scale perception. F πl Input spatial awareness attention module πs The input spatial awareness attention module learns spatial offsets through self-learning. Δ pk and importance scalar Δ mk Feature aggregation is performed on sparse sampling points, enabling the model to focus on key spatial regions of the target, suppressing interference, and obtaining spatially enhanced features. F πs :

[0075] ,in, L Represents the number of feature levels. K It is the number of sparse sampling locations. pk It is a spatial scalar.

[0076] Spatial perception and attention πs Employing a sparsity strategy, through self-learned spatial offsets Δ pk and importance scalar Δ mk By focusing on key locations on the feature map, we can learn which regions in the spatial location of the feature map contain more important information for object detection, thereby enabling feature aggregation on sparse sampling points.

[0077] Step 4, Task-Aware Attention Processing: This involves processing the features perceived through spatial awareness. F πs Input Task Aware Attention Module πc The task-aware attention module adaptively allocates and optimizes feature channels for different sub-tasks such as classification and localization through gating functions, enhancing the decoder's adaptability to multiple tasks and ultimately generating feature maps for object detection. :

[0078] Where α and β are adaptive parameters, and It is a gating function that learns to control the activation threshold. First of all Global average pooling is performed along the dimension, followed by two fully connected layers and a normalization layer. Finally, the sigmoid function is applied to normalize the output. scope ，F πs,c This is the feature slice of the c-th channel. Task-aware attention. πc The activation threshold of each channel is controlled by a gating function (adaptive learning parameter). and Dynamically enhance or suppress specific feature channels to improve the decoder's discriminative power and task adaptability, making it better suited to the tasks or target categories that need to be processed or identified in SAR images.

[0079] Step S3: Train the constructed ArgusSAR model based on the training set, optimize the model parameters through a multi-task joint loss function, and obtain the ArgusSAR model with the best recognition performance.

[0080] During the training phase, the loss function consists of the localization loss. Confidence loss and category loss Composition, defined as follows:

[0081] .

[0082] The CIOU definition formula is as follows:

[0083] ,in, This represents the square of the distance between the center points of the ground truth bounding box and the predicted bounding box, while This represents the diagonal length of the smallest bounding rectangle that can contain both boxes. and The calculation formula is as follows:

[0084] , .

[0085] Based on this, Taking into account overlapping area, center point distance, and aspect ratio consistency, a comprehensive regression objective is provided, the calculation formula of which is defined as follows:

[0086] .

[0087] Confidence loss The definition is as follows: .

[0088] in, This indicates that a target exists. This indicates that there is no target. This represents the balancing weights of the loss caused by negative samples. It is used to reduce the contribution of a large number of negative samples (predicted boxes that do not contain objects) to the loss function, preventing the model from being too inclined to predict everything as background. This represents the bounding box predicted by the model. s' represents the actual bounding box, while s' represents the number of rows and columns of the image.

[0089] The class loss is represented by the following formula: ,in, The category label represents the target object (1 indicates it is the target object, 0 indicates it is not the target object), and This represents the probability predicted by the model.

[0090] Step S4: Input the test set into the ArgusSAR model with the best recognition effect, and use the model to detect ship and aircraft targets in SAR images, and output the location and category recognition results of the target objects.

[0091] To achieve a more comprehensive quantitative evaluation of the performance of the ArgusSAR model, this invention selects the following evaluation metrics: Precision (P), Recall (R), F1 score (F1−score), and Mean Average Precision (mAP). These four different metrics measure the robustness and accuracy of the detector (model) from different perspectives.

[0092] IoU is the ratio of the overlapping region to the combined region of the predicted and ground truth bounding boxes, measuring the degree of overlap between the model's predicted bounding boxes and the ground truth bounding boxes. In this experiment, the IoU is set to 0.5. When the overlap area between the predicted and ground truth boxes exceeds 50%, the detection result is considered correct. The confusion matrix based on IoU is defined in Table 1.

[0093]

[0094] Based on the four types of results (TP, FP, FN, TN) defined in the above confusion matrix, the detection performance of the model is further quantified from different dimensions using the following four core evaluation metrics:

[0095] Accuracy (P): The proportion of truly positive results (correctly detected targets) out of all results predicted as "positive" by the model (i.e., detected targets). The formula is as follows:

[0096] ,

[0097] Recall (R): The proportion of all real-world positive instances (i.e., all objects actually present in the image) that are successfully detected by the model. The formula is as follows:

[0098] ,

[0099] F1 score (F1-score): The harmonic mean of precision (P) and recall (R), a comprehensive evaluation metric designed to balance precision and recall. The calculation formula is as follows:

[0100] ,

[0101] Mean Average Precision ): It is the most core and representative comprehensive evaluation index in the field of object detection, capable of measuring the average detection accuracy of a model across all categories, and it is achieved through... The area under the curve takes into account the model's performance at different confidence thresholds. The higher the value, the better the overall performance of the detector. The calculation formula is as follows:

[0102] .

[0103] The algorithm comparison includes comparison with baseline algorithms, visual comparative analysis, and comparison with improved algorithms from home and abroad.

[0104] First, baseline algorithm comparison: Experiments were conducted on three datasets: SSDD, HRSID, and SADD. The proposed ArgusSAR detector was compared with existing baseline methods. The results are listed in Tables 2, 3, and 4. Relevant detection results are listed below. Figure 5 , Figure 6 and Figure 7 Table 2 shows the comparison results with the baseline algorithm on the SSDD and HRSID ship datasets.

[0105]

[0106] (1) SSDD: As shown in Table 2, on the SSDD dataset, the proposed method shows a 4.9% advantage in F1 score compared to the YOLOv5 algorithm, indicating that it can accurately detect ship targets in complex environmental backgrounds. Compared to the most popular YOLOv8 algorithm, the proposed ArgusSAR increases the detection metrics P, R, and mAP by 1.6%, 2.0%, and 1.5%, respectively, achieving significant performance advantages at the quantitative level. Even compared to the state-of-the-art baseline algorithm YOLOv10, ArgusSAR performs better in ship localization and classification tasks, verifying the effectiveness of ArgusSAR. Furthermore, compared to the SSD algorithm, ArgusSAR improves R and mAP by 12.7% and 10.2%, respectively, demonstrating that the proposed method has excellent detection performance and practical application value.

[0107] (2) HRSID: As shown in Table 2, the backgrounds of images with a large number of small ship objects in the HRSID dataset are more complex. Benefiting from the content-aware upsampling operator, the ArgusSAR detector can accurately reconstruct the ship feature map and distinguish the differences between ship targets and interfering objects, achieving an F1 score of 91.4%, which is 1.9% and 0.4% higher than YOLOv8 and YOLOv10, respectively, effectively balancing the P and R indices. It is worth noting that the performance of the YOLOv7 algorithm degrades significantly when ship targets are densely arranged in SAR images and when targets occlude each other, resulting in more missed detections. Therefore, this experiment can more intuitively demonstrate the excellent performance of ArgusSAR when facing densely arranged ship targets at a quantitative level.

[0108] (3) SADD: Table 3 shows the comparison results with baseline algorithms on the SADD aircraft dataset. As shown in Table 3, the ArgusSAR algorithm proposed in this invention achieves the best performance in both F1 and mAP metrics, quantitatively confirming that ArgusSAR is more suitable for aircraft target detection in SAR imagery compared to mainstream baseline algorithms. Compared with YOLOv8, although ArgusSAR... Slightly lacking in terms of indicators, but in The performance index is 2.9% higher, still demonstrating advanced detection performance. Furthermore, compared to other algorithms, ArgusSAR outperforms all baseline methods in all metrics, proving that ArgusSAR maintains excellent performance against small aircraft objects even in the face of complex interference, making it suitable for SAR imagery-based aircraft detection missions.

[0109]

[0110] Second, visual comparative analysis: Figure 5The results show that the SSD algorithm performs poorly, exhibiting numerous false detections and missing some mutually occluded ship targets (as shown in the red box in the figure). The YOLOv3 and YOLOv4 algorithms reduce false detections compared to SSD, but still miss some difficult-to-detect mutually occluded targets (such as the topmost ship). The YOLOv6 algorithm still exhibits some false detections in this scenario and misses some ship targets (such as the bottommost ship). The detection results of the YOLOv7 to YOLOv10 series algorithms show improved accuracy in ship target detection, but they fail to resolve all false detection issues (such as the topmost ship), indicating that mutually occluded targets pose a significant challenge to detection performance. The ArgusSAR detector proposed in this invention provides particularly accurate and complete detection results in this scenario, successfully detecting almost all ship targets, including instances previously missed by other algorithms, while maintaining an extremely low false detection rate, demonstrating robust detection performance.

[0111] Figure 6 The results show that the SSD algorithm generates a large number of false detections (as shown in the blue box in the figure), with performance degradation being more severe in land areas and coastal boundaries, misdetecting some land equipment as ship targets, while also exhibiting a large number of false negatives (as shown in the red box in the figure), making it almost ineffective for such tasks. While the YOLOv3 and YOLOv4 algorithms are slightly better than SSD, they still have high false detection and false negative rates, and their ability to identify densely packed small ship targets is very limited. YOLOv6 further improves processing capabilities based on YOLOv5, improving false detection and false negative rates, but it still cannot effectively handle sea surface interference. The YOLOv9 and YOLOv10 algorithms can effectively identify most dense small ship targets from high clutter backgrounds, with significantly improved false negative rates and low false detection rates. In comparison, the ArgusSAR detector proposed in this invention demonstrates excellent performance on the HRSID dense small target ship dataset. The detection results control the probability of false detections and effectively detect all real ship targets (almost none of which are in red boxes), proving the strong robustness and excellent efficiency of the ArgusSAR detector in dealing with the challenges of detecting extremely dense and small targets in SAR images.

[0112] Figure 7The results show that the SSD algorithm exhibits a high false detection rate on this dataset (as shown in the blue box in the figure), especially at image edges and in some target gap areas, where it easily misidentifies background clutter or non-target structures as aircraft targets. Compared to SSD, the YOLOv3 and YOLOv4 algorithms have reduced the false detection rate, but they still have shortcomings in background suppression, especially in the right edge area of ​​the image where there are still a large number of false detections. The YOLOv7 and YOLOv9 algorithms both show good detection performance, accurately identifying and locating almost all aircraft targets, achieving high precision and high recall. Meanwhile, the detection performance of the ArgusSAR detector proposed in this invention significantly surpasses that of the advanced YOLO series algorithms, effectively overcoming the phenomenon that baseline algorithms are prone to false detection under background clutter interference, demonstrating high-precision SAR image target detection capabilities. Table 4 shows the comparison results with the improved algorithms on the SSDD ship dataset.

[0113]

[0114] Table 4 shows that the AFSar algorithm has a significant advantage in the R metric. However, it is worth noting that the ArgusSAR algorithm proposed in this invention outperforms the AFSar method by 3.7% in the P metric, by 0.9% in the mAP metric, and by 1.5% in the F1 metric. This indicates that the algorithm of this invention can better balance detection accuracy and recall, and has more robust detection performance. Furthermore, although CSD-YOLO obtained the same mAP metric value as the ArgusSAR algorithm proposed in this invention, the ArgusSAR algorithm has superior quantitative performance in the P and R metrics, exceeding CSD-YOLO by 1.9% and 1.5%, respectively. These quantitative experiments demonstrate that the ArgusSAR algorithm can overcome the effects of mutual occlusion between ships and background interference when facing complex coastal backgrounds, and is more robust in detecting SAR ship targets.

[0115] Table 5 shows that the BL-Net algorithm performs exceptionally well in the R-index, reaching 89.7%. However, the ArgusSAR algorithm proposed in this invention achieves a P-index of 93.5%, tied for the highest with the FESAR algorithm, and also obtains the highest F1 score of 91.4% among all compared methods. This F1 score exceeds that of the BL-Net algorithm by 0.8% and the FESAR algorithm by 1.9%, respectively. Meanwhile, ArgusSAR's R-index of 89.4% is only slightly lower than that of the BL-Net algorithm, fully demonstrating that the ArgusSAR algorithm of this invention can more effectively balance detection accuracy and recall, achieving superior overall detection performance. Furthermore, although the YOLOv5s-MFC algorithm achieves the highest mAP score of 94.5%, the ArgusSAR algorithm proposed in this invention outperforms YOLOv5s-MFC in all three core metrics—P, R, and F1—by 0.3%, 0.3%, and 0.3%, respectively. This further proves that the ArgusSAR algorithm has stronger robustness and more balanced detection capabilities when processing densely packed small target ship datasets such as HRSID. Table 5 shows the comparison results between the improved algorithm and the HRSID ship dataset.

[0116]

[0117] Table 6 shows that the ArgusSAR algorithm proposed in this invention achieves an R-index of 98.6%, the highest among all compared methods, demonstrating its superiority in detecting all real targets. More importantly, ArgusSAR's F1 and mAP indices also significantly outperform all other compared algorithms with excellent scores of 97.8% and 98.9%, respectively. The comparison reveals that although ArgusSAR's P-index (97.1%) is slightly lower than EST-YOLOv5s, its F1 index exceeds EST-YOLOv5s by 2.1%, and its mAP index is also 1.1% higher. This fully demonstrates that the algorithm of this invention can maintain a high level of detection accuracy in scenarios such as SADD aircraft data, where target structural features are complex and geometric information extraction is difficult, and achieves a better balance and robustness in overall performance. Table 6 shows the comparison results with the improved algorithm on the SADD aircraft dataset.

[0118]

[0119] Third, ablation experiments: Tables 7, 8, and 9 present the results for all possible combinations of MHA (for spatial encoders), FFAE (for fine-grained feature aggregation encoders), and MAAD (for multidimensional perceptual attention decoders) on different datasets. A "√" sign indicates that the module is included, while an "×" sign indicates that the module is not included.

[0120] On the SSDD dataset, in experiment A1 (Table 7), we used the MHA module, which improved P, R, and mAP by 2.9%, 4.1%, and 1.7%, respectively. In experiments A4 and A5, the MHA module was combined with the FFAE and MAAD modules, respectively, further improving the algorithm's performance. In experiment A5, when the MHA module was combined with the MAAD module, the mAP increased to 98.8%. Specifically, the feature map processed by the attention module contains a large amount of weighted spatial information, and the MAAD module can enhance the awareness of spatial location by making full use of the information in the feature map. In experiment A6, we did not use the MHA module, which means that the interaction between channels in the feature map cannot be considered, resulting in a decrease in detection performance compared to experiments A4 and A5. Experiments A2, A4, and A6 show that using the FFAE module can significantly improve the R index. This is because the FFAE module considers the influence of other neighboring pixels and uses weighted feature maps to obtain rich detail information and avoids discontinuities in grayscale values ​​after sampling. According to Experiment A7, combining all modules can significantly improve detection performance, with P increasing by 4.6%, R by 5.1%, and mAP by 2.4%. Table 7 shows the ablation experimental results on the SSDD dataset.

[0121]

[0122] In the HRSID dataset, the results in Table 8 show that each module significantly contributes to improving detection performance. In Experiment B1, which only includes the MHA module, the P, R, and mAP metrics improved by 4.5%, 3.4%, and 2.5%, respectively. The significantly improved R metric in Experiment B4 highlights FFAE's ability to capture detailed content information from neighboring pixels. However, in Experiment B6, the lack of sufficient interaction between channel and spatial information due to the absence of the MHA module led to a decrease in detection performance. Experiments B2, B4, and B6 confirm that FFAE, through content-aware upsampling operators, can fuse precise detail information and rich semantic information, thereby significantly improving the R metric. Finally, in Experiment B7, the algorithm achieved optimal performance by using all three modules simultaneously, with P increasing by 5.1%, R by 2.7%, and mAP by 2.9%. Table 8 presents the ablation experiment results on the HRSID dataset.

[0123]

[0124] On the SADD dataset, in Experiment C1 of Table 9, using only the MHA module, the P, R, and mAP were improved by 2.6%, 2.7%, and 4.2%, respectively. In Experiment C4, the combined use of MHA and FFAE significantly improved the P and R metrics, accurately distinguishing the target features from the surrounding environment features, while also demonstrating the superior performance of FFAE in multi-scale feature interaction and fusion.

[0125] Experiments C2, C4, and C6, using the FFAE module, verified the significant impact of the FFAE module on the R-index, demonstrating the detector's superior feature aggregation capability. Table 9 presents the ablation experimental results on the SADD dataset.

[0126]

[0127] Based on this, in order to further confirm the detection capabilities and practical application value of ArgusSAR, Figure 8 A visualized attention heatmap is shown (areas closer to red indicate areas the detector focuses on when it's active). Clearly, the ArgusSAR detector pays more attention to ships and aircraft in SAR imagery, while the YOLOv10 algorithm, in contrast, is more susceptible to background interference. This reflects the ArgusSAR detector's ability to overcome ship occlusion and background interference, making it more robust in detecting typical SAR land and sea targets, thus ensuring the detector focuses on objects of interest rather than interfering structures.

[0128] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes that can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention are all within the protection scope of the claims of the present invention.

Claims

1. A method of monitoring a ship aircraft radar based on Argus SAR, characterized in that, The method comprises the following steps: Step S1, acquiring SAR image data related to ships and aircrafts, performing multi-target labeling and dataset division on the SAR image data, and obtaining a multi-target SAR image dataset containing a training set and a test set; Step S2, constructing an ArgusSAR model through a spatial encoder, a fine-grained feature aggregation encoder, and a multi-dimensional perception attention decoder; the spatial encoder comprises a convolutional encoding module and a multi-head attention module, and is used to extract local interaction and global interaction features of a feature map; the fine-grained feature aggregation encoder comprises a convolution kernel prediction module and a content perception reconstruction module, and is used to realize hierarchical multi-scale feature fusion; the multi-dimensional perception attention decoder comprises scale perception attention, spatial perception attention, and task perception attention, and is used to realize scale perception, spatial perception, and task perception of a target; Step S3, training the constructed ArgusSAR model based on the training set, optimizing model parameters through a multi-task joint loss function, and obtaining an ArgusSAR model with optimal recognition effect; Step S4, inputting the test set into the ArgusSAR model with optimal recognition effect, detecting ship and aircraft targets in the SAR image through the model, and outputting position and category recognition results of the target objects.

2. The method of monitoring a ship aircraft based on Argus SAR according to claim 1, characterized in that, In step S2, the specific working method of the spatial encoder comprises: Step S211, input embedding: input the feature representation as an embedding matrix wherein is the sequence length, is the embedding dimension; Step S212, linear transformation: for each attention head , the embedding matrix is operated with three different weight matrices respectively after input, three linear transformations are performed to obtain a query matrix , a key matrix and a value matrix , and the operation formula is as follows: ， in, , and Is with the first The weight matrix associated with each attention head; Step S213, attention score calculation: when calculating the attention scores of different channels and spatial positions, for each attention head , first calculate the dot product between the query matrix and the key matrix , scaled by the square root of the key dimension , and then apply the Softmax function to obtain the attention score matrix, the operation formula is as follows: ； Step S214, Weighted Summation: Using the attention score matrix Log-value matrix Perform a weighted sum operation to obtain the first... The output of each attention head The calculation formula is as follows: ； Step S215, concatenation transformation: obtain the output of the attention head After obtaining the output of the attention head, all the output matrices are concatenated, and linear transformation is performed on the concatenated output matrix to obtain the final weighted embedding representation The operation formula is as follows: wherein, Wo is an initial random weight matrix.

3. The method of monitoring a ship aircraft based on Argus SAR as claimed in claim 1, wherein, In step S2, the specific working method of the fine-grained feature aggregation encoder comprises: Step S221, according to the content of the input low-resolution feature map, the convolution kernel prediction module dynamically generates specific up-sampling kernels for the subsequent process according to the SAR image content, specifically: Step S2211, input features After channel compression, the features are obtained , and the features are processed by a convolution layer with a convolution kernel size of , and the output channel number is ; σ is the up-sampling ratio, is the content-aware kernel size;​ Step S2212, these spatially spread, for each position of the target high-resolution map an up-sampling kernel of size is generated; Step S2213, independently performing Softmax normalization on each generated up-sampling kernel to ensure that the weight sum is 1, and obtaining the final up-sampling kernel used to guide the content perception reconstruction module; Step S222, for each target position on the output feature map , the convolution kernel prediction module first locates and extracts a local neighborhood with a size of on its corresponding low-resolution input feature map, obtains a specific up-sampling kernel with a size of predicted for the target position; and performs weighted summation on the features in the extracted low-resolution neighborhood using the up-sampling kernel to obtain the final output high-resolution fusion feature map .

4. The method of monitoring a ship aircraft based on Argus SAR as claimed in claim 1, wherein, In step S2, the specific working method of the multi-dimensional perception attention decoder comprises: Step S231, Feature Alignment and Stacking: Receive multi-level feature maps from the fine-grained feature aggregation encoder Align all feature maps by adjusting size and channel number, then stack them into a unified feature tensor Prepare for subsequent processing; Step S232, scale-aware attention processing: input the feature tensor to a scale-aware attention module π l The scale-aware attention module calculates the weights of different feature levels by the following formula, and weights the features to dynamically fuse features of different scales, enhance the adaptability to different size targets, and obtain scale-enhanced features F πl : wherein C represents the number of channels, S represents the number of feature scales, is a linear function, is an activation function; Step S233, spatial-aware attention processing: the scale-aware features F πl are input to the spatial-aware attention module πs , the spatial-aware attention module learns the spatial offset Δ pk and the importance scalar by self-learning Δ mk , and aggregates the features at sparse sampling points, so that the model can focus on the key spatial area of the target and suppress interference to obtain spatially enhanced features F πs : , wherein, L represents the number of characteristic levels, K is the number of sparse sampling locations, pk is a spatial position scalar; Step S234, task-aware attention processing: the spatially perceived features F πs input to the task-aware attention module πc The task-aware attention module adaptively allocates and optimizes feature channels for different sub-tasks through a gating function, enhances the adaptability of the decoder to multi-tasks, and finally generates a feature map for target detection : , wherein, α and β are adaptive parameters, and is a gating function that learns the activation threshold, first global average pooling is performed in the dimension, then two fully connected layers and a normalization layer are used, and finally a Sigmoid function is applied to normalize the output to the range; F πs,c is the feature slice of the c-th channel.

5. The method of monitoring a ship aircraft based on Argus SAR as claimed in claim 1, wherein, In step S3, the multi-task joint loss function is composed of a positioning loss , a confidence loss and a category loss , defined as follows: wherein, (1) Positioning loss The calculation formula is: ； wherein, represents the square of the distance between the center points of the real box and the predicted box, while represents the diagonal length of the minimum circumscribed rectangle that can simultaneously contain the two boxes, the parameters and The calculation formula is as follows: , ; IOU is the intersection over union of the predicted box and the real box, ω gt is the width of the real box, is the width of the predicted box, h gt is the height of the real box, h is the height of the predicted box; (2) Confidence loss The calculation formula is: wherein, represents the presence of a target, represents the absence of a target, represents a balancing weight for the loss caused by negative samples to reduce the contribution of a large number of negative samples to the loss function, represents a bounding box predicted by the model, represents a real bounding box, while s' represents the number of rows and columns of the picture; (3) Category loss The calculation formula is: ， wherein, a class label representing a target object, representing a probability predicted by the model.

Images